Robust rotation-invariant fingerprint image hashing using fast wavelet transform and fractal coding for noise filtering with minimal run-time Authors: Lakshmi Dathan, Anjaneyulu G.S.G.N Published: 2026 ABSTRACT Biometric data, particularly fingerprints, provide a reliable means of personal identification; however, protecting sensitive biometric templates against unauthorized access, tampering, and reconstruction remains a significant challenge in biometric security systems. Existing fingerprint hashing approaches often face limitations in terms of computational efficiency, robustness to geometric distortions, and preservation of discriminative ridge features. To address these limitations, this article proposes a fingerprint hashing technique that integrates the Fast Wavelet Transform (FWT), Fourier–Mellin Transform (FMT), and fractal coding to improve robustness and computational efficiency. In the proposed framework, FWT is employed for noise reduction and multi-resolution feature extraction, effectively preserving essential ridge structures and improving image quality. The FMT is utilized to achieve rotation-invariant feature representation, while fractal coding performs efficient image compression and enhances robustness against distortions and dimensional variations. The integration of FWT with fractal coding further reduces computational complexity while maintaining strong discriminative capability in the generated hash codes. Comprehensive experiments conducted on the FVC2002, FVC2004, and SOCOFing fingerprint databases demonstrate the effectiveness of the proposed hashing framework. The method achieves strong discriminability with an Equal Error Rate (EER) as low as 0.3864%, while maintaining low False Match Rate (FMR) and False Non-Match Rate (FNMR). In addition, the proposed approach significantly reduces computational cost, achieving an average execution time of 2.507 seconds compared to 1048.888 seconds in existing fractal-coding-based hashing approaches. Performance evaluation using PSNR, SNR, SSIM, entropy, mutual information, and edge preservation metrics confirms that the proposed framework improves noise resilience, preserves structural fingerprint features, and maintains robustness against geometric variations. These results demonstrate that the proposed approach generates compact, secure, and robust fingerprint hashes suitable for biometric authentication, forensic identification, and digital identity verification systems. Introduction Biometric authentication systems based on fingerprints, facial recognition, and iris patterns have become integral to modern security frameworks due to their ability to leverage unique physiological traits for reliable identity verification. Unlike conventional password-based systems, biometric data is inherently distinctive and difficult to replicate1, enabling robust and statistically consistent authentication under controlled conditions. Recent integration with deep learning techniques has further enhanced performance, particularly in complex recognition scenarios such as unconstrained facial identification, thereby expanding the applicability of biometric systems in security-critical environments. Despite these advancements, biometric systems introduce significant security and privacy challenges. Since biometric traits are non-revocable, their compromise leads to permanent identity risks. Common threats include spoofing attacks, where fabricated biometric samples are used to deceive authentication systems2, and replay attacks involving the reuse of previously captured data. To address these vulnerabilities, modern systems incorporate liveness detection, multi-factor authentication, encryption, and biometric template protection mechanisms to safeguard stored data and prevent reconstruction or misuse3. Among biometric modalities, fingerprints remain one of the most reliable due to their permanence and distinctiveness, as ridge patterns are formed during fetal development and remain stable throughout life4. Fingerprint Image Hashing provides a compact and fixed-length representation of fingerprint data, preserving essential structural information while enabling efficient storage and matching. Unlike traditional hashing, it is specifically designed to retain discriminative biometric characteristics. The hashing pipeline typically consists of preprocessing, feature extraction, and secure hash generation. The development of fingerprint hashing has evolved from early feature-based methods, such as minutiae extraction and ridge analysis, to more advanced indexing techniques including locality-sensitive hashing (LSH). More recently, deep learning approaches based on convolutional neural networks (CNNs)5 and generative adversarial networks (GANs) have demonstrated strong feature learning capabilities. However, these approaches often face challenges related to robustness against noise, geometric variations (rotation and scaling), and high computational complexity in large-scale deployments. Although transform-domain approaches such as wavelet-based methods6, Fourier–Mellin invariance7, and fractal coding techniques8 have improved robustness and representation efficiency, several limitations persist. In particular, fractal-based hashing methods suffer from high computational cost due to exhaustive domain–range matching, limited resilience to non-Gaussian noise, and a predominant focus on recognition accuracy rather than security-centric hash design. These limitations hinder their practical deployment in real-time, large-scale, and security-sensitive applications. Recent studies have explored hybrid transform-domain representations and lightweight encoding strategies to balance robustness and computational efficiency9. While these methods demonstrate improved performance, they often lack a unified design that simultaneously addresses computational efficiency, noise robustness, and security-oriented hashing requirements. To bridge these gaps, this work proposes a unified fingerprint hashing framework with the following key contributions: FWT-guided fractal coding for computational efficiency: the proposed method introduces a wavelet-guided constraint on fractal domain–range matching, significantly reducing search complexity while preserving self-similarity properties critical for secure hashing.Noise-adaptive multi-resolution hashing: a multi-resolution Fast Wavelet Transform (FWT) filtering strategy is employed to suppress noise across frequency bands, enhancing robustness against non-Gaussian noise (e.g., salt-and-pepper and speckle) while preserving ridge structures.Security-oriented hash design: unlike recognition-driven approaches, the framework explicitly targets biometric template protection by ensuring fixed-length representation, high entropy10, collision resistance, revocability, and robustness to content-preserving transformations.The novelty of the proposed work lies not in the individual components—fast wavelet transform (FWT), Fourier–Mellin Transform (FMT), and fractal coding—but in their structured and interdependent integration within a unified hashing pipeline. Specifically, wavelet-domain coefficients are leveraged not only for noise suppression but also to guide fractal encoding, thereby reducing computational overhead while preserving discriminative structural information. Combined with FMT-based normalization for rotation and scale invariance, this coordinated design achieves an effective balance between robustness, efficiency, and security. FWT-guided fractal coding for computational efficiency: the proposed method introduces a wavelet-guided constraint on fractal domain–range matching, significantly reducing search complexity while preserving self-similarity properties critical for secure hashing. FWT-guided fractal coding for computational efficiency: the proposed method introduces a wavelet-guided constraint on fractal domain–range matching, significantly reducing search complexity while preserving self-similarity properties critical for secure hashing. Noise-adaptive multi-resolution hashing: a multi-resolution Fast Wavelet Transform (FWT) filtering strategy is employed to suppress noise across frequency bands, enhancing robustness against non-Gaussian noise (e.g., salt-and-pepper and speckle) while preserving ridge structures. Noise-adaptive multi-resolution hashing: a multi-resolution Fast Wavelet Transform (FWT) filtering strategy is employed to suppress noise across frequency bands, enhancing robustness against non-Gaussian noise (e.g., salt-and-pepper and speckle) while preserving ridge structures. Security-oriented hash design: unlike recognition-driven approaches, the framework explicitly targets biometric template protection by ensuring fixed-length representation, high entropy10, collision resistance, revocability, and robustness to content-preserving transformations. Security-oriented hash design: unlike recognition-driven approaches, the framework explicitly targets biometric template protection by ensuring fixed-length representation, high entropy10, collision resistance, revocability, and robustness to content-preserving transformations. From a broader perspective, transform-domain feature extraction techniques play a central role in fingerprint representation. Wavelet transforms capture localized ridge details through multi-resolution analysis, while FMT provides inherent invariance to geometric transformations. Orthogonal polynomial descriptors, such as Zernike and Legendre moments, offer compact global representations but are often sensitive to noise and segmentation errors. The proposed framework exploits the complementary strengths of wavelet analysis and FMT to achieve robust and efficient hashing, while avoiding the limitations associated with purely global descriptors. Recent advances in orthogonal polynomial-based methods11–13 further highlight the effectiveness of transform-domain representations, though they often require careful parameter tuning. The remainder of the paper is organized as follows. “Literature review” reviews existing fingerprint hashing approaches. “Motivation and objectives” presents the motivation and contributions. “Basic infrastructure and biometric authentication view” describes the mathematical foundations. “Proposed hashing scheme” details the proposed framework, followed by experimental evaluation in “Experimental results” and efficiency analysis in “Efficiency analysis”. Security and privacy analysis are presented in“Security and privacy analysis”, and conclusions with future directions are discussed in “Conclusion and future scope”. Literature review Fingerprint authentication is a widely adopted biometric security mechanism, and fingerprint hashing has become an essential approach for secure template representation. Over the years, several methods have been proposed to generate robust, discriminative, and compact hash representations while ensuring computational efficiency and robustness to distortions. Capelli et al.14 introduced a fingerprint identification scheme based on Locality-Sensitive Hashing (LSH) and Minutiae Cylinder-Code (MCC), where minutiae features are mapped into fixed-length transformation-invariant binary vectors for efficient large-scale search. Although the method improves indexing speed, its performance depends heavily on accurate minutiae extraction, making it sensitive to noisy or low-quality fingerprints. Wang et al.15 proposed a shape-context-based hashing approach using SIFT-Harris feature detection to enhance robustness against content-preserving distortions. Extending this idea, Muthu et al.16 applied SIFT-Harris features with shape context descriptors for fingerprint hashing. While these approaches offer strong robustness to geometric transformations and compression, they suffer from high computational cost due to complex feature extraction. Jin et al.17 presented a kernel-learning-based fingerprint representation that converts minutiae features into fixed-length binary codes. Despite its effectiveness, kernel selection and binarization introduce computational overhead and potential quantization errors. Similarly, Wang et al.18 proposed a cancelable fingerprint template using zoned minutiae pairs and discrete Fourier transform (DFT), which enhances template security but depends on accurate zone definition and may expose reconstruction vulnerabilities. Kho et al.19 developed a method that encodes minutiae structures into ordered fixed-length bit strings using elliptical Gaussian modeling and Euclidean normalization, improving matching efficiency but limiting compatibility with conventional systems. Abdullahi et al.20 combined minutiae features with shape context descriptors to generate compact hash signatures; however, the method incurs increased computational overhead and lacks extensive validation across diverse datasets. Hybrid approaches have also been explored to improve robustness and security. Shahzad et al.21 proposed an alignment-free cancelable fingerprint template using partial discrete wavelet transform (PDWT) and a window-shift-XOR model, achieving improved attack resistance at the cost of higher complexity. Wang et al.22 introduced a hybrid perceptual hashing framework combining global, local, and structural features, which enhances tampering detection but increases computational requirements. Birouk et al.23 integrated SIFT, Harris filtering, discrete wavelet transform (DWT), and singular value decomposition (SVD) to generate robust binary hash codes, though the multi-stage pipeline reduces scalability. Gupta et al.24 combined fingerprint recognition with SHA-512 for enhanced security, but with significant computational cost. Hammad et al.25 proposed a cancelable multi-biometric framework using non-negative matrix factorization (NMF) and lightweight deep learning models, which improves accuracy and privacy but depends on training data generalization. Deep learning-based approaches have further advanced fingerprint representation. Anand et al.26 introduced PoreNet, a CNN-based model for pore-level feature learning, while Chen et al.5 proposed a hybrid Siamese CNN combined with SIFT descriptors for improved robustness under varying conditions. Vallabhadas et al.27 also developed a cancelable CNN-based template protection scheme. Despite their effectiveness, these methods require large datasets, high computational resources, and complex training procedures, limiting their applicability in lightweight systems. Recent studies, including28, emphasize the need for fingerprint hashing frameworks that balance robustness, computational efficiency, and template security. Overall, existing methods face challenges such as sensitivity to noise and minutiae extraction errors, high computational complexity, or dependence on large training datasets. To provide a clearer comparison of representative fingerprint hashing techniques, Table 1 summarizes several existing approaches along with their advantages and limitations.Table 1Summary of representative fingerprint hashing methods.StudyMethodAdvantagesLimitationsCapelli et al.14MCC + LSHFast indexing and scalable searchSensitive to minutiae extraction errorsWang et al.15Shape context hashingRobust to geometric transformationsHigh computational complexityAbdullahi et al.20Minutiae + shape contextCompact hash representationComputational overheadBirouk et al.23SIFT + DWT + SVDRobust feature representationMultiple transformations increase complexityGupta et al.24Fingerprint + SHA-512Strong cryptographic protectionHigh computational requirementsDeep learning methods5,26,27CNN-based descriptorsStrong representation learningRequires large datasets and high computation Table 1Summary of representative fingerprint hashing methods.StudyMethodAdvantagesLimitationsCapelli et al.14MCC + LSHFast indexing and scalable searchSensitive to minutiae extraction errorsWang et al.15Shape context hashingRobust to geometric transformationsHigh computational complexityAbdullahi et al.20Minutiae + shape contextCompact hash representationComputational overheadBirouk et al.23SIFT + DWT + SVDRobust feature representationMultiple transformations increase complexityGupta et al.24Fingerprint + SHA-512Strong cryptographic protectionHigh computational requirementsDeep learning methods5,26,27CNN-based descriptorsStrong representation learningRequires large datasets and high computation Summary of representative fingerprint hashing methods. From the above studies, it can be observed that many existing fingerprint hashing approaches face challenges such as high computational complexity, sensitivity to noise and minutiae extraction errors, or reliance on large training datasets. To address these limitations, this study proposes a fingerprint hashing framework that integrates the Fast Wavelet Transform (FWT), Fourier–Mellin Transform (FMT), and fractal coding. The combination of these techniques enables effective noise suppression, rotation-invariant feature extraction, and compact fingerprint representation while maintaining computational efficiency and robustness without relying on learning-based models. Motivation and objectives In 2020, Abdullahi et al. proposed “Fractal coding-based robust and alignment-free fingerprint image hashing”29, which employs the Fourier-Mellin Transform (FMT) and fractal coding to achieve rotation-invariant fingerprint hashing. The FMT effectively handles rotation variations, ensuring hash consistency regardless of fingerprint orientation. Additionally, Gaussian filtering and Haar wavelet decomposition are used in preprocessing to enhance image quality prior to hashing. This article builds upon and improves the approach in29 by addressing several inherent limitations. Although fractal coding effectively captures self-similar structures, it is computationally expensive due to exhaustive domain–range block matching, making it less suitable for large datasets. Furthermore, Gaussian filtering, while effective for noise suppression, tends to blur ridge structures and minutiae, which are critical for fingerprint recognition. It is also primarily suited for Gaussian noise and may not handle non-Gaussian noise effectively. To overcome these issues, the proposed method introduces the Fast Wavelet Transform (FWT) as an efficient preprocessing technique. Unlike Gaussian filtering, FWT provides multi-resolution analysis that suppresses noise while preserving important fingerprint features such as ridges and minutiae. Incorporating FWT prior to fractal encoding not only improves feature robustness but also reduces computational complexity by limiting unnecessary domain–range comparisons. To justify these improvements, the discriminability of the existing method29 is first evaluated. Table 2 presents the False Non-Match Rate (FNMR)30 and False Match Rate (FMR)30 under various manipulations, serving as a benchmark for comparison.Table 2FNMR and FMR of the existing algorithm29.ManipulationsFNMR (%)FMR (%)Gamma correction2.191.25Gaussian filtering1.391.69JPEG compression0.000.00Median filtering5.582.64Rotation0.000.00Scaling0.000.00 Table 2FNMR and FMR of the existing algorithm29.ManipulationsFNMR (%)FMR (%)Gamma correction2.191.25Gaussian filtering1.391.69JPEG compression0.000.00Median filtering5.582.64Rotation0.000.00Scaling0.000.00 FNMR and FMR of the existing algorithm29. The primary objective of this work is to develop a robust, secure, and rotation-invariant fingerprint hashing framework integrating FWT, FMT, and fractal coding. The specific objectives are:Enhancing fingerprint security and discriminability by effectively distinguishing genuine and imposter fingerprints while maintaining robustness to noise and distortions.Achieving rotation and scale invariance using Fourier–Mellin transform (FMT).Improving noise filtering through fast wavelet transform (FWT) while preserving minutiae and ridge structures.Optimizing computational efficiency by integrating FWT with fractal compression.Experimentally validating the method on real datasets and assessing robustness against content-preserving manipulations. Enhancing fingerprint security and discriminability by effectively distinguishing genuine and imposter fingerprints while maintaining robustness to noise and distortions. Enhancing fingerprint security and discriminability by effectively distinguishing genuine and imposter fingerprints while maintaining robustness to noise and distortions. Achieving rotation and scale invariance using Fourier–Mellin transform (FMT). Achieving rotation and scale invariance using Fourier–Mellin transform (FMT). Improving noise filtering through fast wavelet transform (FWT) while preserving minutiae and ridge structures. Improving noise filtering through fast wavelet transform (FWT) while preserving minutiae and ridge structures. Optimizing computational efficiency by integrating FWT with fractal compression. Optimizing computational efficiency by integrating FWT with fractal compression. Experimentally validating the method on real datasets and assessing robustness against content-preserving manipulations. Experimentally validating the method on real datasets and assessing robustness against content-preserving manipulations. Basic infrastructure and biometric authentication view Fast wavelet transform (FWT) Fast wavelet transform (FWT) is an efficient technique widely used in image processing for multi-resolution analysis and noise suppression. It improves computational efficiency by enabling the fast computation of Discrete Wavelet Transform (DWT) coefficients. In practical signal and image processing applications, the DWT is efficiently implemented using the fast wavelet transform (FWT) based on a multiresolution filter-bank structure6. The transform decomposes a signal into a set of wavelet coefficients that represent localized frequency information at different scales. This multiresolution representation makes wavelet transforms particularly suitable for analyzing image structures such as edges and textures. For a discrete-time signal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x[n]$$\end{document}, the wavelet coefficients can be expressed as1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} c(a,b)=\frac{1}{\sqrt{a}}\sum _{n=-\infty }^{\infty } x[n]^{*}\psi \left( \frac{n-b}{a}\right) \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} c(a,b)&= Wavelet \ coefficient \ at \ {{\bf a}} \ and \ {{\bf b}}, \ \ \psi (t) = Wavelet \ function \\ a&= Scale \ parameter , \ \ \ b = Translation \ parameter \end{aligned}$$\end{document}FWT is implemented by a filter bank structure that comprises stages of filtering and downsampling. Generally speaking, the steps involved in the decomposition procedure are as follows: Filtering: the input signal is passed through a low-pass filter to separate its low-frequency components, called Approximate coefficients. On the other hand, the signal is passed through a high-pass filter to extract the high-frequency components called Detailed coefficients. These filters are very meticulously designed to work in such a manner as to match their frequency responses with the scaling function and the wavelet function used in the DWT. 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} c[n] = h[n]^{*}x[n], \ \ \ \ d[n] = g[n]^{*}x[n] \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} c[n]&= Approximate \ coefficient, \ \ \ \ d[n] = Detailed \ coefficient\\ h[n]&= Low-pass \ filter, \ \ \ \ g[n] = High-pass \ filter \end{aligned}$$\end{document}Downsampling: the approximate and detailed coefficients are downsampled to reduce the computational and storage demands. Downsampling gives a low-resolution representation of low-frequency (approximate) and high-frequency (detailed) elements of the signal. 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} c[\frac{n}{2}] = c[n], \ \ \ \ d[\frac{n}{2}] = d[n] \end{aligned}$$\end{document} Through downsampling, redundant information is eliminated, hence enabling efficient analysis and processing of the signal at different scales. The downsampled coefficients can be decomposed further to obtain even coarser representations, hence providing a multi-scale analysis of the signal.The downsampled approximate coefficient c[\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{n}{2}$$\end{document}] is then used as the input for the next level of decomposition. The process of filtering, downsampling, and recursion is repeated until the desired level of decomposition is reached. Filtering: the input signal is passed through a low-pass filter to separate its low-frequency components, called Approximate coefficients. On the other hand, the signal is passed through a high-pass filter to extract the high-frequency components called Detailed coefficients. These filters are very meticulously designed to work in such a manner as to match their frequency responses with the scaling function and the wavelet function used in the DWT. 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} c[n] = h[n]^{*}x[n], \ \ \ \ d[n] = g[n]^{*}x[n] \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} c[n]&= Approximate \ coefficient, \ \ \ \ d[n] = Detailed \ coefficient\\ h[n]&= Low-pass \ filter, \ \ \ \ g[n] = High-pass \ filter \end{aligned}$$\end{document} Filtering: the input signal is passed through a low-pass filter to separate its low-frequency components, called Approximate coefficients. On the other hand, the signal is passed through a high-pass filter to extract the high-frequency components called Detailed coefficients. These filters are very meticulously designed to work in such a manner as to match their frequency responses with the scaling function and the wavelet function used in the DWT. 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} c[n] = h[n]^{*}x[n], \ \ \ \ d[n] = g[n]^{*}x[n] \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} c[n]&= Approximate \ coefficient, \ \ \ \ d[n] = Detailed \ coefficient\\ h[n]&= Low-pass \ filter, \ \ \ \ g[n] = High-pass \ filter \end{aligned}$$\end{document} Downsampling: the approximate and detailed coefficients are downsampled to reduce the computational and storage demands. Downsampling gives a low-resolution representation of low-frequency (approximate) and high-frequency (detailed) elements of the signal. 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} c[\frac{n}{2}] = c[n], \ \ \ \ d[\frac{n}{2}] = d[n] \end{aligned}$$\end{document} Through downsampling, redundant information is eliminated, hence enabling efficient analysis and processing of the signal at different scales. The downsampled coefficients can be decomposed further to obtain even coarser representations, hence providing a multi-scale analysis of the signal. Downsampling: the approximate and detailed coefficients are downsampled to reduce the computational and storage demands. Downsampling gives a low-resolution representation of low-frequency (approximate) and high-frequency (detailed) elements of the signal. 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} c[\frac{n}{2}] = c[n], \ \ \ \ d[\frac{n}{2}] = d[n] \end{aligned}$$\end{document} Through downsampling, redundant information is eliminated, hence enabling efficient analysis and processing of the signal at different scales. The downsampled coefficients can be decomposed further to obtain even coarser representations, hence providing a multi-scale analysis of the signal. The downsampled approximate coefficient c[\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{n}{2}$$\end{document}] is then used as the input for the next level of decomposition. The process of filtering, downsampling, and recursion is repeated until the desired level of decomposition is reached. The downsampled approximate coefficient c[\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{n}{2}$$\end{document}] is then used as the input for the next level of decomposition. The process of filtering, downsampling, and recursion is repeated until the desired level of decomposition is reached. Fourier–Mellin transform (FMT) The Fourier–Mellin transform (FMT) is a very powerful technique used in image processing for achieving rotation, scale, and translation invariance. The transformation confirms that fingerprint matching remains invariant under the change in orientation or position. It combines Fourier Transform and Mellin Transform to achieve the above properties. FMT consists of three main steps; Fourier transform: the input image is converted into the frequency domain by applying 2D Fourier Transform31. 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} F(u,v) = \mathcal {F}\{I_{m}(x,y)\} = \sum _{x=0}^{M-1} \sum _{y=0}^{N-1} I_{m}(x,y) e^{-i 2\pi \left( \frac{ux}{M} + \frac{vy}{N} \right) } \end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_{m}(x,y)$$\end{document} represents the input fingerprint image in the spatial domain, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y$$\end{document} denote the spatial pixel coordinates, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N$$\end{document} represent the height and width of the image respectively. The variables \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$u$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v$$\end{document} denote the frequency-domain coordinates, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F(u,v)$$\end{document} represents the Fourier transformed image, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i$$\end{document} denotes the imaginary unit and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {F}\{\cdot \}$$\end{document} represents the two-dimensional Fourier transform operator.Conversion to log-polar coordinates: after conversion into the frequency domain, F(u,v) is converted into log-polar coordinates. This helps in achieving rotation and scale invariance by mapping the frequency components into a coordinate system that gets converted into shifts. The advantage of this transformation is that rotation of the input image shifts along an angle while scaling corresponds according to a factor, that allows more efficient processing of transformations in fingerprint images. This behavior emulates the Mellin Transform, allowing both rotation and scaling to be treated as translations.Final Fourier transform: a 2D Fourier transform is then applied to the log-polar representation. Since rotation and scaling have been transformed into shifts, this final Fourier Transform converts them into phase shifts, enabling efficient correlation and matching. Thus, the combined transformation ensures that fingerprint matching remains robust against changes in orientation, size, or position. Fourier transform: the input image is converted into the frequency domain by applying 2D Fourier Transform31. 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} F(u,v) = \mathcal {F}\{I_{m}(x,y)\} = \sum _{x=0}^{M-1} \sum _{y=0}^{N-1} I_{m}(x,y) e^{-i 2\pi \left( \frac{ux}{M} + \frac{vy}{N} \right) } \end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_{m}(x,y)$$\end{document} represents the input fingerprint image in the spatial domain, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y$$\end{document} denote the spatial pixel coordinates, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N$$\end{document} represent the height and width of the image respectively. The variables \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$u$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v$$\end{document} denote the frequency-domain coordinates, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F(u,v)$$\end{document} represents the Fourier transformed image, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i$$\end{document} denotes the imaginary unit and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {F}\{\cdot \}$$\end{document} represents the two-dimensional Fourier transform operator. Fourier transform: the input image is converted into the frequency domain by applying 2D Fourier Transform31. 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} F(u,v) = \mathcal {F}\{I_{m}(x,y)\} = \sum _{x=0}^{M-1} \sum _{y=0}^{N-1} I_{m}(x,y) e^{-i 2\pi \left( \frac{ux}{M} + \frac{vy}{N} \right) } \end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_{m}(x,y)$$\end{document} represents the input fingerprint image in the spatial domain, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y$$\end{document} denote the spatial pixel coordinates, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N$$\end{document} represent the height and width of the image respectively. The variables \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$u$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v$$\end{document} denote the frequency-domain coordinates, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F(u,v)$$\end{document} represents the Fourier transformed image, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i$$\end{document} denotes the imaginary unit and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {F}\{\cdot \}$$\end{document} represents the two-dimensional Fourier transform operator. Conversion to log-polar coordinates: after conversion into the frequency domain, F(u,v) is converted into log-polar coordinates. This helps in achieving rotation and scale invariance by mapping the frequency components into a coordinate system that gets converted into shifts. The advantage of this transformation is that rotation of the input image shifts along an angle while scaling corresponds according to a factor, that allows more efficient processing of transformations in fingerprint images. This behavior emulates the Mellin Transform, allowing both rotation and scaling to be treated as translations. Conversion to log-polar coordinates: after conversion into the frequency domain, F(u,v) is converted into log-polar coordinates. This helps in achieving rotation and scale invariance by mapping the frequency components into a coordinate system that gets converted into shifts. The advantage of this transformation is that rotation of the input image shifts along an angle while scaling corresponds according to a factor, that allows more efficient processing of transformations in fingerprint images. This behavior emulates the Mellin Transform, allowing both rotation and scaling to be treated as translations. Final Fourier transform: a 2D Fourier transform is then applied to the log-polar representation. Since rotation and scaling have been transformed into shifts, this final Fourier Transform converts them into phase shifts, enabling efficient correlation and matching. Thus, the combined transformation ensures that fingerprint matching remains robust against changes in orientation, size, or position. Final Fourier transform: a 2D Fourier transform is then applied to the log-polar representation. Since rotation and scaling have been transformed into shifts, this final Fourier Transform converts them into phase shifts, enabling efficient correlation and matching. Thus, the combined transformation ensures that fingerprint matching remains robust against changes in orientation, size, or position. Fractal coding Fractal coding is a method of data compression based on the property of self-similarity within the image. This technique exhibits a larger image as a collection of smaller, self-similar patterns by splitting it into smaller range and domain blocks. Each range block is contracted to a corresponding domain block. This process is repeated for all range blocks of the image. The primary objective is to represent the entire image as a set of affine transformations \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T_p$$\end{document} that comprises scaling rotational and translational parameters. The technique identifies domain blocks that best match range blocks under these affine transformations. The parameters are then saved instead of pixel values. Authentication framework In a hash-based fingerprint authentication system, the process begins with a registration phase where the user’s fingerprint is scanned by the biometric scanner. The scanned fingerprint image undergoes a hashing procedure, finally generating a fixed-length compact Hash Digest. This hash acts as the digital template of the fingerprint. Conventionally, both the fingerprint image and the corresponding hash could be retained in the database with the user’s data. However, for improved security, it is best to only store the hash. By excluding the raw fingerprint image from the storage, the system ensures that even in the event of a database breach, attackers cannot reconstruct the original biometric data, thereby preserving user privacy and the system’s integrity. In the authentication step, the user provides his/her fingerprint in the authentication zone. Upon scanning the fingerprint, a hash is created using the same hashing procedure previously used and then compared to the database hashes. The comparison is performed using similarity measures such as Hamming distance32, correlation, or a threshold-based matching approach. If the generated hash is sufficiently similar to a hash in the database within the defined thresholds, the user is authenticated successfully. It is important that the system is collision-resistant with a zero collision rate, which ensures that no two different fingerprint inputs produce the same hash output. This property safeguards against impersonation and enhances the system’s trustworthiness. Figure 1 shows a schematic representation of the authentication system.Fig. 1Authentication framework using fingerprint data via image hashing. Authentication framework using fingerprint data via image hashing. Proposed hashing scheme The proposed fingerprint hashing algorithm follows a structured multi-stage framework to ensure robustness, efficiency, and compact representation. Initially, the input fingerprint image I undergoes preprocessing to enhance quality and suppress acquisition artifacts. The Fast Wavelet Transform (FWT) is then applied to obtain multi-resolution frequency subbands, enabling noise suppression while preserving salient ridge structures. To achieve rotation invariance, the Fourier–Mellin Transform (FMT) is applied to the wavelet-domain representation, mapping the image into a rotation-invariant feature space. Unlike conventional approaches where stages operate independently, the proposed framework introduces interaction between wavelet decomposition and fractal coding, where wavelet coefficients guide the fractal domain–range matching process by restricting the candidate search space based on coefficient similarity. This reduces computational complexity while preserving discriminative information. Finally, fractal coding captures self-similar patterns to produce a compact and stable fingerprint hash. The overall workflow is illustrated in Fig. 2, and detailed steps are provided in Algorithm 1.Fig. 2Block diagram of the proposed fingerprint hashing framework. Block diagram of the proposed fingerprint hashing framework. Algorithm 1Proposed fingerprint hash generation framework. Proposed fingerprint hash generation framework. Pre-processing Pre-processing is an extremely crucial stage in fingerprint image processing where the input image I goes through four pre-processing steps including Normalization, Enhancement, Binarization and Thinning. This is defined as the first step before the next stages set in the hashing framework from “Pre-processing” to “Final hash generation”. Normalization: normalizing pixel values improves contrast and reduces variations in lighting conditions. This algorithm employs the mean-variance normalization algorithm. 5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} normalized \ pixel \ value = \frac{(pixel \ value - mean)}{standard \ deviation} \end{aligned}$$\end{document}Enhancement: this is done to enhance fingerprint features such as ridges and valleys. Filtering, sharpening, and contrast stretching are the main steps involved in this process. Filtering is done to clean the image from artifacts, sharpening to bring out the edge of ridges and valleys, and contrast stretching expands the dynamic range of pixel values, enhancing feature visibility. Histogram equalization method is used here for enhancement; 6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} cdf(c) = \frac{cumulative \ sum(histogram(c)}{total \ pixels} \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} equlized \ pixel = cdf(pixel) \end{aligned}$$\end{document}Binarization: this is often needed before feature extraction and subsequent matching algorithms. 7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} binarized \ pixel = {\left\{ \begin{array}{ll} 1, & \text {if } pixel>threshold \\ 0, & \text {otherwise} \end{array}\right. } \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} threshold = \frac{pixel_{max}-pixel_{min}}{2} \end{aligned}$$\end{document}Thinning: thinning involves iteration, a connectivity rule, and a boundary removal condition. This preserves minutiae points and is essential for precise feature extraction. Morphological thinning is used in this algorithm. Normalization: normalizing pixel values improves contrast and reduces variations in lighting conditions. This algorithm employs the mean-variance normalization algorithm. 5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} normalized \ pixel \ value = \frac{(pixel \ value - mean)}{standard \ deviation} \end{aligned}$$\end{document} Normalization: normalizing pixel values improves contrast and reduces variations in lighting conditions. This algorithm employs the mean-variance normalization algorithm. 5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} normalized \ pixel \ value = \frac{(pixel \ value - mean)}{standard \ deviation} \end{aligned}$$\end{document} Enhancement: this is done to enhance fingerprint features such as ridges and valleys. Filtering, sharpening, and contrast stretching are the main steps involved in this process. Filtering is done to clean the image from artifacts, sharpening to bring out the edge of ridges and valleys, and contrast stretching expands the dynamic range of pixel values, enhancing feature visibility. Histogram equalization method is used here for enhancement; 6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} cdf(c) = \frac{cumulative \ sum(histogram(c)}{total \ pixels} \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} equlized \ pixel = cdf(pixel) \end{aligned}$$\end{document} Enhancement: this is done to enhance fingerprint features such as ridges and valleys. Filtering, sharpening, and contrast stretching are the main steps involved in this process. Filtering is done to clean the image from artifacts, sharpening to bring out the edge of ridges and valleys, and contrast stretching expands the dynamic range of pixel values, enhancing feature visibility. Histogram equalization method is used here for enhancement; 6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} cdf(c) = \frac{cumulative \ sum(histogram(c)}{total \ pixels} \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} equlized \ pixel = cdf(pixel) \end{aligned}$$\end{document} Binarization: this is often needed before feature extraction and subsequent matching algorithms. 7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} binarized \ pixel = {\left\{ \begin{array}{ll} 1, & \text {if } pixel>threshold \\ 0, & \text {otherwise} \end{array}\right. } \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} threshold = \frac{pixel_{max}-pixel_{min}}{2} \end{aligned}$$\end{document} Binarization: this is often needed before feature extraction and subsequent matching algorithms. 7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} binarized \ pixel = {\left\{ \begin{array}{ll} 1, & \text {if } pixel>threshold \\ 0, & \text {otherwise} \end{array}\right. } \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} threshold = \frac{pixel_{max}-pixel_{min}}{2} \end{aligned}$$\end{document} Thinning: thinning involves iteration, a connectivity rule, and a boundary removal condition. This preserves minutiae points and is essential for precise feature extraction. Morphological thinning is used in this algorithm. Thinning: thinning involves iteration, a connectivity rule, and a boundary removal condition. This preserves minutiae points and is essential for precise feature extraction. Morphological thinning is used in this algorithm. Noise removal and minutiae extraction—FWT The pre-processed image I\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{p}$$\end{document} processed from “Pre-processing” is subjected to FWT in order to remove noise and extract minutiae points. This involves the following steps;DecompositionNoise removalMinutiae extraction and filteringThese fundamental steps involved in this procedure are explained in detail below, providing a comprehensive understanding of each phase. Decomposition: I\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{p}$$\end{document} is decomposed into a series of wavelet coefficients at different scales and orientations (Eq. (1)) and analyzes the coefficients to identify noise and minutiae features.Noise removal: a threshold T is set as in the Eq. (8). This type of thresholding gives a greater resolvability to very small values of noise coefficients than to signal coefficients. Discarding coefficients that are less than this threshold and reconstructing the image with the rest can efficiently remove noise while retaining all the significant features of the image. 8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} T = \mu _c+2\sigma _c \end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu _c$$\end{document} is the mean and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma _c$$\end{document} is the standard deviation of wavelet coefficients. For l levels of decomposition, the final reconstruction of the denoised image using the coefficients c’ is given by; 9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} c'(a,b) = {\left\{ \begin{array}{ll} c(a,b), & \text {if } |c(a,b)| \ge T\\ 0, & \text {otherwise} \end{array}\right. } \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} I'_p = \sum (a)^{l-1} \sum (b)^{2^a-1} c'(a,b)^{*} \psi (\frac{n-b}{a}) \end{aligned}$$\end{document}Minutiae extraction: potential minutiae regions can be identified from the detail coefficients through the following steps;Ridge and valley segmentation: apply the Sobel operator on the detail coefficients at all scales. This will compute gradients in the horizontal and vertical directions. Sobel operator33 is defined as; 10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} g_{x} = [[-1, -2, -1], [0, 0, 0], [1, 2, 1]] \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} g_{y} = [[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]] \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} G(a,b) = sqrt(g_{x}*c'(a,b)^{2} + g_{y}*c'(a,b)^{2}) \end{aligned}$$\end{document}Minutiae localization: examine the detail coefficients in the segments to find the minutiae points and categorize them by their structure. 11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} Ridge \ Ending \implies G(a,b) = 1 \ and \sum (i, j \in N) g(i, j) = 1 \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} Bifurcation \implies G(a,b) = 1 \ and \sum (i, j \in N) g(i, j) \ge 3 \end{aligned}$$\end{document}Minutiae filtering: remove minutiae that are too close to each other or to the image boundary, resulting in the final minutiae extracted image I\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{m}$$\end{document}. Decomposition Decomposition Noise removal Noise removal Minutiae extraction and filtering Minutiae extraction and filtering Decomposition: I\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{p}$$\end{document} is decomposed into a series of wavelet coefficients at different scales and orientations (Eq. (1)) and analyzes the coefficients to identify noise and minutiae features. Decomposition: I\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{p}$$\end{document} is decomposed into a series of wavelet coefficients at different scales and orientations (Eq. (1)) and analyzes the coefficients to identify noise and minutiae features. Noise removal: a threshold T is set as in the Eq. (8). This type of thresholding gives a greater resolvability to very small values of noise coefficients than to signal coefficients. Discarding coefficients that are less than this threshold and reconstructing the image with the rest can efficiently remove noise while retaining all the significant features of the image. 8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} T = \mu _c+2\sigma _c \end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu _c$$\end{document} is the mean and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma _c$$\end{document} is the standard deviation of wavelet coefficients. For l levels of decomposition, the final reconstruction of the denoised image using the coefficients c’ is given by; 9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} c'(a,b) = {\left\{ \begin{array}{ll} c(a,b), & \text {if } |c(a,b)| \ge T\\ 0, & \text {otherwise} \end{array}\right. } \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} I'_p = \sum (a)^{l-1} \sum (b)^{2^a-1} c'(a,b)^{*} \psi (\frac{n-b}{a}) \end{aligned}$$\end{document} Noise removal: a threshold T is set as in the Eq. (8). This type of thresholding gives a greater resolvability to very small values of noise coefficients than to signal coefficients. Discarding coefficients that are less than this threshold and reconstructing the image with the rest can efficiently remove noise while retaining all the significant features of the image. 8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} T = \mu _c+2\sigma _c \end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu _c$$\end{document} is the mean and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma _c$$\end{document} is the standard deviation of wavelet coefficients. For l levels of decomposition, the final reconstruction of the denoised image using the coefficients c’ is given by; 9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} c'(a,b) = {\left\{ \begin{array}{ll} c(a,b), & \text {if } |c(a,b)| \ge T\\ 0, & \text {otherwise} \end{array}\right. } \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} I'_p = \sum (a)^{l-1} \sum (b)^{2^a-1} c'(a,b)^{*} \psi (\frac{n-b}{a}) \end{aligned}$$\end{document} Minutiae extraction: potential minutiae regions can be identified from the detail coefficients through the following steps;Ridge and valley segmentation: apply the Sobel operator on the detail coefficients at all scales. This will compute gradients in the horizontal and vertical directions. Sobel operator33 is defined as; 10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} g_{x} = [[-1, -2, -1], [0, 0, 0], [1, 2, 1]] \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} g_{y} = [[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]] \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} G(a,b) = sqrt(g_{x}*c'(a,b)^{2} + g_{y}*c'(a,b)^{2}) \end{aligned}$$\end{document}Minutiae localization: examine the detail coefficients in the segments to find the minutiae points and categorize them by their structure. 11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} Ridge \ Ending \implies G(a,b) = 1 \ and \sum (i, j \in N) g(i, j) = 1 \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} Bifurcation \implies G(a,b) = 1 \ and \sum (i, j \in N) g(i, j) \ge 3 \end{aligned}$$\end{document}Minutiae filtering: remove minutiae that are too close to each other or to the image boundary, resulting in the final minutiae extracted image I\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{m}$$\end{document}. Minutiae extraction: potential minutiae regions can be identified from the detail coefficients through the following steps;Ridge and valley segmentation: apply the Sobel operator on the detail coefficients at all scales. This will compute gradients in the horizontal and vertical directions. Sobel operator33 is defined as; 10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} g_{x} = [[-1, -2, -1], [0, 0, 0], [1, 2, 1]] \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} g_{y} = [[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]] \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} G(a,b) = sqrt(g_{x}*c'(a,b)^{2} + g_{y}*c'(a,b)^{2}) \end{aligned}$$\end{document}Minutiae localization: examine the detail coefficients in the segments to find the minutiae points and categorize them by their structure. 11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} Ridge \ Ending \implies G(a,b) = 1 \ and \sum (i, j \in N) g(i, j) = 1 \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} Bifurcation \implies G(a,b) = 1 \ and \sum (i, j \in N) g(i, j) \ge 3 \end{aligned}$$\end{document}Minutiae filtering: remove minutiae that are too close to each other or to the image boundary, resulting in the final minutiae extracted image I\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{m}$$\end{document}. Ridge and valley segmentation: apply the Sobel operator on the detail coefficients at all scales. This will compute gradients in the horizontal and vertical directions. Sobel operator33 is defined as; 10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} g_{x} = [[-1, -2, -1], [0, 0, 0], [1, 2, 1]] \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} g_{y} = [[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]] \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} G(a,b) = sqrt(g_{x}*c'(a,b)^{2} + g_{y}*c'(a,b)^{2}) \end{aligned}$$\end{document} Ridge and valley segmentation: apply the Sobel operator on the detail coefficients at all scales. This will compute gradients in the horizontal and vertical directions. Sobel operator33 is defined as; 10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} g_{x} = [[-1, -2, -1], [0, 0, 0], [1, 2, 1]] \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} g_{y} = [[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]] \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} G(a,b) = sqrt(g_{x}*c'(a,b)^{2} + g_{y}*c'(a,b)^{2}) \end{aligned}$$\end{document} Minutiae localization: examine the detail coefficients in the segments to find the minutiae points and categorize them by their structure. 11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} Ridge \ Ending \implies G(a,b) = 1 \ and \sum (i, j \in N) g(i, j) = 1 \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} Bifurcation \implies G(a,b) = 1 \ and \sum (i, j \in N) g(i, j) \ge 3 \end{aligned}$$\end{document} Minutiae localization: examine the detail coefficients in the segments to find the minutiae points and categorize them by their structure. 11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} Ridge \ Ending \implies G(a,b) = 1 \ and \sum (i, j \in N) g(i, j) = 1 \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} Bifurcation \implies G(a,b) = 1 \ and \sum (i, j \in N) g(i, j) \ge 3 \end{aligned}$$\end{document} Minutiae filtering: remove minutiae that are too close to each other or to the image boundary, resulting in the final minutiae extracted image I\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{m}$$\end{document}. Minutiae filtering: remove minutiae that are too close to each other or to the image boundary, resulting in the final minutiae extracted image I\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{m}$$\end{document}. Rotation, scaling and translation invariance—Fourier–Mellin transform In this section of the algorithm, the fingerprint image is transformed into the frequency domain using Fourier transform (FFT) and then a reference minutia is chosen. The remaining minutiae are converted into polar coordinates, aligned, and mapped onto a fixed-length polar grid so that a binary bit string can be extracted from it, representing the fingerprint in a compact form. This section is derived from29. However, for greater clarity and flow, a detailed explanation of the steps involved is provided to enhance the reader’s understanding. Transitioning to the frequency domain: The final minutiae extracted image I\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{m}$$\end{document} obtained from “Noise removal and minutiae extraction—FWT” is transformed into the frequency domain using the 2D Fourier Transform (FFT) (Eq. 4). The magnitude spectrum of the transformed image using Eq. (12) provides insight into the spatial frequency characteristics of the fingerprint, allowing the analysis and identification of important features such as ridges, valleys, and noise patterns. 12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} |F(u,v)| = \log (1+|F(u,v)|) \end{aligned}$$\end{document}Selection of reference minutia and conversion to polar coordinates: a reference minutia \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M_r = (x_r, y_r)$$\end{document} is selected from the set of detected minutiae points to serve as the origin for the polar coordinate transformation. In this work, the reference minutia is chosen as the minutia point that lies closest to the centroid of the detected minutiae set. The centroid \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(x_c, y_c)$$\end{document} of the minutiae set is computed by averaging the coordinates of all detected minutiae points. The minutia with minimum Euclidean distance from this centroid is selected as the reference minutia. This selection strategy ensures spatial stability and reduces sensitivity to boundary distortions, providing a consistent reference point for the subsequent polar coordinate transformation. Once the reference minutia is selected, all other minutiae points \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M_i=(x_i,y_i)$$\end{document} are converted into polar coordinates relative to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M_r$$\end{document}. Equation (13) gives the polar conversion of each \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M_i$$\end{document} for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i=1,2,3,\dots$$\end{document}. This transformation makes the system translation invariant, since absolute positions of minutiae are now represented in a relative manner. 13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \rho _i = \sqrt{(x_i-x_r)^2+(y_i-y_r)^2} \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \alpha _i = \tan ^{-1}\left( \frac{y_i-y_r}{x_i-x_r}\right) \frac{180}{\pi } \end{aligned}$$\end{document} Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _i$$\end{document} represents the distance of the minutiae point from the reference minutia and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha _i$$\end{document} represents the angular orientation of the minutia relative to the reference point.Minutiae alignment: the following transformation is applied in order to align the minutiae points with the reference orientation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _r$$\end{document}. This ensures rotation invariance, making the system robust to rotational or orientational changes in the fingerprint image. 14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \begin{bmatrix} x_i^t \\ y_i^t \end{bmatrix} = \begin{bmatrix} \cos \theta _r & -\sin \theta _r \\ \sin \theta _r & \cos \theta _r \end{bmatrix} \begin{bmatrix} x_i - x_r \\ y_i - y_r \end{bmatrix} + \begin{bmatrix} x_r \\ y_r \end{bmatrix} \end{aligned}$$\end{document}Polar representation: A polar image grid \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_g(\rho ,\phi )$$\end{document} is generated by mapping the aligned minutiae \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(x_i^t,y_i^t)$$\end{document} onto a circular grid. This representation facilitates scale invariance and improves robustness to variations in fingerprint image size. The transformation with R being the radius of the polar grid is given by 15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \phi _i=\left\lfloor \frac{( \alpha _i + 180 ) \times R}{360} \right\rfloor \end{aligned}$$\end{document}Minutiae bit string generation: Polar segments are generated from the polar grid by dividing it into angular sectors. Each sector is examined for the presence of minutiae points to encode the binary string \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_m$$\end{document}, which forms the binary representation of the fingerprint features. The encoding rule is defined as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \text {Bit}(j) = {\left\{ \begin{array}{ll} 1, & \text {if more than one minutia exists in the segment} \\ 0, & \text {otherwise} \end{array}\right. } \end{aligned}$$\end{document}Second Fourier transform: A second 2D Fourier transform (FFT) is applied to the polar grid image \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_g(\rho ,\phi )$$\end{document}, resulting in the transformed image \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_f$$\end{document}. This additional transformation further enhances invariance to rotation, scaling, and translation, improving the robustness of the generated fingerprint hash. Transitioning to the frequency domain: The final minutiae extracted image I\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{m}$$\end{document} obtained from “Noise removal and minutiae extraction—FWT” is transformed into the frequency domain using the 2D Fourier Transform (FFT) (Eq. 4). The magnitude spectrum of the transformed image using Eq. (12) provides insight into the spatial frequency characteristics of the fingerprint, allowing the analysis and identification of important features such as ridges, valleys, and noise patterns. 12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} |F(u,v)| = \log (1+|F(u,v)|) \end{aligned}$$\end{document} Transitioning to the frequency domain: The final minutiae extracted image I\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{m}$$\end{document} obtained from “Noise removal and minutiae extraction—FWT” is transformed into the frequency domain using the 2D Fourier Transform (FFT) (Eq. 4). The magnitude spectrum of the transformed image using Eq. (12) provides insight into the spatial frequency characteristics of the fingerprint, allowing the analysis and identification of important features such as ridges, valleys, and noise patterns. 12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} |F(u,v)| = \log (1+|F(u,v)|) \end{aligned}$$\end{document} Selection of reference minutia and conversion to polar coordinates: a reference minutia \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M_r = (x_r, y_r)$$\end{document} is selected from the set of detected minutiae points to serve as the origin for the polar coordinate transformation. In this work, the reference minutia is chosen as the minutia point that lies closest to the centroid of the detected minutiae set. The centroid \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(x_c, y_c)$$\end{document} of the minutiae set is computed by averaging the coordinates of all detected minutiae points. The minutia with minimum Euclidean distance from this centroid is selected as the reference minutia. This selection strategy ensures spatial stability and reduces sensitivity to boundary distortions, providing a consistent reference point for the subsequent polar coordinate transformation. Once the reference minutia is selected, all other minutiae points \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M_i=(x_i,y_i)$$\end{document} are converted into polar coordinates relative to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M_r$$\end{document}. Equation (13) gives the polar conversion of each \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M_i$$\end{document} for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i=1,2,3,\dots$$\end{document}. This transformation makes the system translation invariant, since absolute positions of minutiae are now represented in a relative manner. 13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \rho _i = \sqrt{(x_i-x_r)^2+(y_i-y_r)^2} \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \alpha _i = \tan ^{-1}\left( \frac{y_i-y_r}{x_i-x_r}\right) \frac{180}{\pi } \end{aligned}$$\end{document} Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _i$$\end{document} represents the distance of the minutiae point from the reference minutia and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha _i$$\end{document} represents the angular orientation of the minutia relative to the reference point. Selection of reference minutia and conversion to polar coordinates: a reference minutia \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M_r = (x_r, y_r)$$\end{document} is selected from the set of detected minutiae points to serve as the origin for the polar coordinate transformation. In this work, the reference minutia is chosen as the minutia point that lies closest to the centroid of the detected minutiae set. The centroid \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(x_c, y_c)$$\end{document} of the minutiae set is computed by averaging the coordinates of all detected minutiae points. The minutia with minimum Euclidean distance from this centroid is selected as the reference minutia. This selection strategy ensures spatial stability and reduces sensitivity to boundary distortions, providing a consistent reference point for the subsequent polar coordinate transformation. Once the reference minutia is selected, all other minutiae points \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M_i=(x_i,y_i)$$\end{document} are converted into polar coordinates relative to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M_r$$\end{document}. Equation (13) gives the polar conversion of each \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$M_i$$\end{document} for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i=1,2,3,\dots$$\end{document}. This transformation makes the system translation invariant, since absolute positions of minutiae are now represented in a relative manner. 13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \rho _i = \sqrt{(x_i-x_r)^2+(y_i-y_r)^2} \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \alpha _i = \tan ^{-1}\left( \frac{y_i-y_r}{x_i-x_r}\right) \frac{180}{\pi } \end{aligned}$$\end{document} Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _i$$\end{document} represents the distance of the minutiae point from the reference minutia and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha _i$$\end{document} represents the angular orientation of the minutia relative to the reference point. Minutiae alignment: the following transformation is applied in order to align the minutiae points with the reference orientation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _r$$\end{document}. This ensures rotation invariance, making the system robust to rotational or orientational changes in the fingerprint image. 14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \begin{bmatrix} x_i^t \\ y_i^t \end{bmatrix} = \begin{bmatrix} \cos \theta _r & -\sin \theta _r \\ \sin \theta _r & \cos \theta _r \end{bmatrix} \begin{bmatrix} x_i - x_r \\ y_i - y_r \end{bmatrix} + \begin{bmatrix} x_r \\ y_r \end{bmatrix} \end{aligned}$$\end{document} Minutiae alignment: the following transformation is applied in order to align the minutiae points with the reference orientation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _r$$\end{document}. This ensures rotation invariance, making the system robust to rotational or orientational changes in the fingerprint image. 14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \begin{bmatrix} x_i^t \\ y_i^t \end{bmatrix} = \begin{bmatrix} \cos \theta _r & -\sin \theta _r \\ \sin \theta _r & \cos \theta _r \end{bmatrix} \begin{bmatrix} x_i - x_r \\ y_i - y_r \end{bmatrix} + \begin{bmatrix} x_r \\ y_r \end{bmatrix} \end{aligned}$$\end{document} Polar representation: A polar image grid \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_g(\rho ,\phi )$$\end{document} is generated by mapping the aligned minutiae \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(x_i^t,y_i^t)$$\end{document} onto a circular grid. This representation facilitates scale invariance and improves robustness to variations in fingerprint image size. The transformation with R being the radius of the polar grid is given by 15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \phi _i=\left\lfloor \frac{( \alpha _i + 180 ) \times R}{360} \right\rfloor \end{aligned}$$\end{document} Polar representation: A polar image grid \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_g(\rho ,\phi )$$\end{document} is generated by mapping the aligned minutiae \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(x_i^t,y_i^t)$$\end{document} onto a circular grid. This representation facilitates scale invariance and improves robustness to variations in fingerprint image size. The transformation with R being the radius of the polar grid is given by 15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \phi _i=\left\lfloor \frac{( \alpha _i + 180 ) \times R}{360} \right\rfloor \end{aligned}$$\end{document} Minutiae bit string generation: Polar segments are generated from the polar grid by dividing it into angular sectors. Each sector is examined for the presence of minutiae points to encode the binary string \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_m$$\end{document}, which forms the binary representation of the fingerprint features. The encoding rule is defined as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \text {Bit}(j) = {\left\{ \begin{array}{ll} 1, & \text {if more than one minutia exists in the segment} \\ 0, & \text {otherwise} \end{array}\right. } \end{aligned}$$\end{document} Minutiae bit string generation: Polar segments are generated from the polar grid by dividing it into angular sectors. Each sector is examined for the presence of minutiae points to encode the binary string \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_m$$\end{document}, which forms the binary representation of the fingerprint features. The encoding rule is defined as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \text {Bit}(j) = {\left\{ \begin{array}{ll} 1, & \text {if more than one minutia exists in the segment} \\ 0, & \text {otherwise} \end{array}\right. } \end{aligned}$$\end{document} Second Fourier transform: A second 2D Fourier transform (FFT) is applied to the polar grid image \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_g(\rho ,\phi )$$\end{document}, resulting in the transformed image \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_f$$\end{document}. This additional transformation further enhances invariance to rotation, scaling, and translation, improving the robustness of the generated fingerprint hash. Second Fourier transform: A second 2D Fourier transform (FFT) is applied to the polar grid image \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_g(\rho ,\phi )$$\end{document}, resulting in the transformed image \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_f$$\end{document}. This additional transformation further enhances invariance to rotation, scaling, and translation, improving the robustness of the generated fingerprint hash. Image compression-fractal coding Fractal coding has numerous advantages including high compression ratios. However, The computational complexity of this compression technique is really high. To combat this, the system have employed the Fast Wavelet Transform (FWT) to optimize efficiency. Here, the range and domain blocks are matched based on the coefficient values. The range block is matched with the domain block with the closest coefficient value and this reduces the time complexity. The following are the steps involved in this section.Image partitioningFeature extraction using FWTFractal matching and extractionFractal element quantizationA comprehensive explanation of each of these steps is elaborated in detail to ensure clarity and to highlight how FWT contributes to reducing computational overhead while preserving critical image features. Image partitioning: the transformed image \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_f$$\end{document} generated from “Rotation, scaling and translation invariance—Fourier–Mellin transform” into blocks of size VxV, from which Non-overlapping Range Blocks of size nxn are generated. Overlapping Domain Blocks of size 2nx2n is divided from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_f$$\end{document}. The number of range (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R_n$$\end{document}) and domain blocks (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D_n$$\end{document}) is calculated as in29. The choice of block size directly influences both the compression efficiency and feature stability of the fractal coding process. Smaller blocks provide finer structural representation but increase computational complexity, whereas larger blocks reduce computation but may lose local ridge information.FWT for feature extraction: applying wavelet transform to both range and domain blocks generates the corresponding wavelet detailed and approximate coefficients. These coefficients are stored for the next matching process.Fractal matching and extraction: the range block is matched to the best suitable domain block by comparing the detailed coefficients. The position of the matched domain \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(B_jx,B_jy)$$\end{document} is extracted and the feature vector is generated. Two feature vectors \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(FR_i,FB_j)$$\end{document} are generated using the detailed coefficients of range and domain blocks. The position and feature vectors together form the fractal elements.Fractal element quantization: the fractal elements are then quantized. Feature vectors are quantized by rounding and taking the absolute values of the coefficients. The quantized fractal elements are then stored as a compact feature vector. Image partitioning Image partitioning Feature extraction using FWT Feature extraction using FWT Fractal matching and extraction Fractal matching and extraction Fractal element quantization Fractal element quantization Image partitioning: the transformed image \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_f$$\end{document} generated from “Rotation, scaling and translation invariance—Fourier–Mellin transform” into blocks of size VxV, from which Non-overlapping Range Blocks of size nxn are generated. Overlapping Domain Blocks of size 2nx2n is divided from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_f$$\end{document}. The number of range (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R_n$$\end{document}) and domain blocks (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D_n$$\end{document}) is calculated as in29. The choice of block size directly influences both the compression efficiency and feature stability of the fractal coding process. Smaller blocks provide finer structural representation but increase computational complexity, whereas larger blocks reduce computation but may lose local ridge information. Image partitioning: the transformed image \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_f$$\end{document} generated from “Rotation, scaling and translation invariance—Fourier–Mellin transform” into blocks of size VxV, from which Non-overlapping Range Blocks of size nxn are generated. Overlapping Domain Blocks of size 2nx2n is divided from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_f$$\end{document}. The number of range (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R_n$$\end{document}) and domain blocks (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D_n$$\end{document}) is calculated as in29. The choice of block size directly influences both the compression efficiency and feature stability of the fractal coding process. Smaller blocks provide finer structural representation but increase computational complexity, whereas larger blocks reduce computation but may lose local ridge information. FWT for feature extraction: applying wavelet transform to both range and domain blocks generates the corresponding wavelet detailed and approximate coefficients. These coefficients are stored for the next matching process. FWT for feature extraction: applying wavelet transform to both range and domain blocks generates the corresponding wavelet detailed and approximate coefficients. These coefficients are stored for the next matching process. Fractal matching and extraction: the range block is matched to the best suitable domain block by comparing the detailed coefficients. The position of the matched domain \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(B_jx,B_jy)$$\end{document} is extracted and the feature vector is generated. Two feature vectors \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(FR_i,FB_j)$$\end{document} are generated using the detailed coefficients of range and domain blocks. The position and feature vectors together form the fractal elements. Fractal matching and extraction: the range block is matched to the best suitable domain block by comparing the detailed coefficients. The position of the matched domain \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(B_jx,B_jy)$$\end{document} is extracted and the feature vector is generated. Two feature vectors \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(FR_i,FB_j)$$\end{document} are generated using the detailed coefficients of range and domain blocks. The position and feature vectors together form the fractal elements. Fractal element quantization: the fractal elements are then quantized. Feature vectors are quantized by rounding and taking the absolute values of the coefficients. The quantized fractal elements are then stored as a compact feature vector. Fractal element quantization: the fractal elements are then quantized. Feature vectors are quantized by rounding and taking the absolute values of the coefficients. The quantized fractal elements are then stored as a compact feature vector. Final hash generation Hash digest generation: the compact feature vector processed from “Image compression-fractal coding” is converted into a binary hash \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_q$$\end{document}. This is concatenated with the Minutiae Bit String \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H_m$$\end{document} which resulted from “Noise removal and minutiae extraction—FWT” to generate the Final Hash Digest H for the input Image I.16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} H=H_m||H_q \end{aligned}$$\end{document} Experimental results Various experiments are performed to analyze the effectiveness of the proposed fingerprint image hashing algorithm using multiple publicly available fingerprint databases. Specifically, experiments are conducted on the standard FVC2002, FVC2004 (DB1, DB2, DB3, and DB4)34 and SOCOFing35 fingerprint datasets. Each FVC database contains 10 distinct classes with 8 impressions per class. For consistency and to minimize intra-class noise variations, the clearest impression among the eight available samples is selected for each fingerprint, resulting in a total of 80 representative images from the FVC datasets. The SOCOFing dataset is incorporated to evaluate the generalization capability of the proposed hashing framework on a larger and more diverse dataset. From the SOCOFing database, which contains multiple impressions per fingerprint, the clearest impression among ten samples for each fingerprint is selected, resulting in a total of 600 high-quality images, which together with the FVC datasets yielded 680 fingerprint images used for experimental analysis. Figure 3 shows two sample images from each of these databases. The proposed fingerprint hashing framework was implemented in MATLAB R2024a36. The grayscale fingerprint images are resized to 512\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}512 pixels and underwent morphological thinning to enhance the ridge structure. The fast wavelet transform (FWT) was performed using the Haar wavelet with three levels of decomposition to capture multi-scale fingerprint ridge information. The Fourier–Mellin transform (FMT) was implemented with the radial coordinate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho$$\end{document} restricted to M/2, and the angular space was divided into 36 segments, each covering \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^\circ$$\end{document}. For fractal encoding, the fingerprint image was partitioned into non-overlapping range blocks of size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$8 \times 8$$\end{document} and overlapping domain blocks of size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$16 \times 16$$\end{document}. The total number of range (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R_n$$\end{document}) and domain blocks(\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D_n$$\end{document}) is computed as;17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} R_n = \frac{V^2}{n^2}=\frac{64^2}{8^2}=64 \end{aligned}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} D_n = (V-2n+1)^2 = (64-16+1)^2=2401 \end{aligned}$$\end{document}The domain blocks were downsampled during matching to maintain scale consistency with range blocks. Each resulting coefficient was quantized and encoded using an 8-bit representation, resulting in a fixed-length 4096-bit fingerprint hash. All experiments were conducted on an HP 250 G8 notebook computer equipped with an 11th Generation Intel Core i3-1115G4 processor (3.00 GHz) and 8 GB RAM. The implementation does not require specialized hardware or GPU acceleration, demonstrating the practical feasibility of the proposed hashing framework on standard computing platforms.Fig. 3Representative fingerprint images from the datasets used in this study: FVC2002 (DB1–DB4), FVC2004 (DB1–DB4), and the SOCOFing dataset. Two sample images are shown for each database for illustration purposes. Representative fingerprint images from the datasets used in this study: FVC2002 (DB1–DB4), FVC2004 (DB1–DB4), and the SOCOFing dataset. Two sample images are shown for each database for illustration purposes. The sample outputs obtained for each step of the algorithm are as follows (Figs. 4, 5, 6, 7, 8, 9); Step 1: Pre-processing Step 2: Noise removal and minutiae extraction—FWT Step 3: Rotation, scaling and translation invariance—Fourier–Mellin transform Step 4: Image compression-fractal coding Step 5: Final hash generation Fig. 4Sample image for each pre-processing steps. Fig. 5Wavelet coefficient from FWT. Fig. 6Sample denoised image and minutiae points. Fig. 7Rotation, scaling and translation invariant image representations. Fig. 8Sample range blocks, domain blocks, and fractal elements. Fig. 9Final hash digest. Sample image for each pre-processing steps. Wavelet coefficient from FWT. Sample denoised image and minutiae points. Rotation, scaling and translation invariant image representations. Sample range blocks, domain blocks, and fractal elements. Final hash digest. While the configuration mentioned earlier is used as the default setting for all experiments, an additional parameter sensitivity analysis is conducted to evaluate the robustness of the proposed hashing framework with respect to key design parameters. This is performed by systematically varying key design parameters while keeping the remaining settings fixed. Specifically, the wavelet decomposition level, fractal block size, and number of polar angular segments are varied across practical ranges. The analysis is conducted using multiple impressions of the same fingerprint, and the resulting intra-class Hamming distances are evaluated in terms of mean and standard deviation. Hamming distance is widely used in perceptual hashing systems because the generated hash codes are represented as binary bit strings. It measures the number of bit positions at which two hashes differ, providing a simple yet effective similarity measure between fingerprint templates. Due to its low computational complexity and direct applicability to binary hash representations, Hamming distance is particularly suitable for large-scale biometric authentication systems where rapid matching and efficient storage are essential. Table 3 summarizes the quantitative results, while Fig. 10 illustrates the impact of parameter variation on hash stability.Table 3Summary of parameter sensitivity analysis of the proposed fingerprint hashing framework.Parameter variedRangeObserved trendWavelet decomposition level1–3Minor variation, stable HDBlock size4–16No observable impactPolar angular segments18–72Gradual increase in HD Table 3Summary of parameter sensitivity analysis of the proposed fingerprint hashing framework.Parameter variedRangeObserved trendWavelet decomposition level1–3Minor variation, stable HDBlock size4–16No observable impactPolar angular segments18–72Gradual increase in HD Summary of parameter sensitivity analysis of the proposed fingerprint hashing framework. Fig. 10Parameter sensitivity on hash stability. Parameter sensitivity on hash stability. The parameter sensitivity analysis demonstrates that the proposed fingerprint hashing framework exhibits strong robustness with respect to variations in key design parameters. Across all tested wavelet decomposition levels and block sizes, the intra-class Hamming distance remains nearly unchanged, indicating that the extracted hash representation is stable and not sensitive to moderate structural variations in the transform domain. In contrast, increasing the number of polar angular segments results in a gradual rise in the Hamming distance, which is expected due to finer angular resolution capturing more directional details rather than indicating instability. Importantly, no abrupt performance degradation is observed for any parameter setting. These results confirm that the proposed hashing scheme does not rely on carefully tuned parameters and maintains consistent behavior across a wide operating range, thereby reinforcing its robustness and practical applicability in real-world biometric systems. Efficiency analysis The robustness of the fingerprint hashing algorithm is evaluated by the Hamming distance computed between original and content-preserving manipulated fingerprint images. Discriminability is analyzed by evaluating the False Match Rate (FMR), False Non-Match Rate (FNMR), and Equal Error Rate (EER)30 across genuine and impostor fingerprints. The noise resilience of the algorithm is further tested by introducing salt-and-pepper and speckle noise. The Signal-to-Noise Ratio (SNR) and Peak Signal-to-Noise Ratio (PSNR)37 values are computed to compare Fast Wavelet Transform (FWT) and Gaussian filtering. Further assessment included edge preservation and Structural Similarity Index (SSIM)38 to establish that the integrity of fingerprint features is preserved. The computational efficacy of the proposed algorithm and29 is also evaluated. Robustness analysis against content-preserving manipulations The robustness of the proposed fingerprint hashing method is evaluated by measuring the Hamming distance between the original fingerprint hash and hashes generated from content-preserving manipulated images. Six different types of manipulations are considered, namely gamma correction, Gaussian filtering, median filtering, scaling, JPEG compression, and rotation (pixel-wise). Each manipulation type is applied using three different parameter values, as shown in Table 4. These manipulations represent common image processing operations that may occur during fingerprint acquisition, transmission, or storage. A threshold of 800 bits is used for decision making, such that a Hamming distance below this threshold indicates that the manipulated fingerprint remains content-preserving and therefore corresponds to the same original fingerprint. Across the FVC2002, FVC2004, and SOCOFing datasets, a total of 680 original fingerprint samples are considered. For each original image, six manipulation types are applied using three parameter values, resulting in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$680 \times 6 \times 3 = 12,240$$\end{document} manipulated fingerprint images used for robustness evaluation. Figure 11 illustrates the probability density distribution of normalized Hamming distances for perceptually similar (original–manipulated) and dissimilar (different–manipulated) fingerprint pairs. The distribution corresponding to original–manipulated pairs is concentrated at lower normalized Hamming distance values, with its mean located around 0.12. In contrast, the distribution for different–manipulated pairs is centered at significantly higher values, with a mean close to 0.6. This clear separation between the two distributions demonstrates that manipulated fingerprints remain much closer to their corresponding originals than to fingerprints from different subjects. Such a wide margin between genuine and imposter distributions is essential for robust fingerprint verification and indicates strong discriminability of the proposed hashing framework. The manipulation ranges shown in Table 4 are selected according to commonly used evaluation settings in perceptual hashing and biometric robustness studies reported in the literature. These ranges represent realistic variations that may occur during practical fingerprint acquisition and processing.Table 4Manipulation parameters used for robustness testing.Types of manipulations (6)ParameterNo. of manipulations consideredSet 1 valueSet 2 valueSet 3 valueGamma correction\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma$$\end{document}0.90.951.05Gaussian filter\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma$$\end{document}0.40.50.95JPEG compressionQuality304050Median filterFilter Size\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2\times 2$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times 3$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5\times 5$$\end{document}Rotation (pixel-wise)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$15^\circ$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$20^\circ$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$25^\circ$$\end{document}ScalingScale Factor0.80.91.05 Table 4Manipulation parameters used for robustness testing.Types of manipulations (6)ParameterNo. of manipulations consideredSet 1 valueSet 2 valueSet 3 valueGamma correction\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma$$\end{document}0.90.951.05Gaussian filter\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma$$\end{document}0.40.50.95JPEG compressionQuality304050Median filterFilter Size\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2\times 2$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times 3$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5\times 5$$\end{document}Rotation (pixel-wise)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$15^\circ$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$20^\circ$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$25^\circ$$\end{document}ScalingScale Factor0.80.91.05 Manipulation parameters used for robustness testing. Fig. 11Normalized hamming distance distribution. Normalized hamming distance distribution. Remark 1 A scaling value of 1 does not introduce any change to the original image. Therefore, the practical range for scaling is considered between 0.8 and 1.05 (excluding 1). Similar considerations apply to gamma correction and median filtering operations. Remark 2 The parameter ranges summarized in Table 4 represent the acceptable manipulation levels under which the fingerprint hash remains stable and can still be matched with the original fingerprint image. Remark 3 When the manipulation parameters exceed the specified ranges shown in Table 5, the resulting fingerprint hash deviates significantly from the original hash and the fingerprint is correctly identified as a different sample. Table 5Range of parameters for manipulations.Types of manipulations (6)ParameterRangeGamma correction \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma$$\end{document} 0.9–1.05 (except 1)Gaussian filter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma$$\end{document} 0.4–0.95JPEG compression Quality 30–40Median filter Filter Size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2\times 2$$\end{document}–\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5\times 5$$\end{document}Rotation (pixel-wise) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$15^\circ$$\end{document}–\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$25^\circ$$\end{document}Scaling Scale Factor 0.8–1.05 (except 1) Table 5Range of parameters for manipulations.Types of manipulations (6)ParameterRangeGamma correction \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma$$\end{document} 0.9–1.05 (except 1)Gaussian filter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma$$\end{document} 0.4–0.95JPEG compression Quality 30–40Median filter Filter Size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2\times 2$$\end{document}–\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5\times 5$$\end{document}Rotation (pixel-wise) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$15^\circ$$\end{document}–\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$25^\circ$$\end{document}Scaling Scale Factor 0.8–1.05 (except 1) Range of parameters for manipulations. In addition to the predefined manipulation ranges, extended experiments were conducted using more severe distortion levels outside the recommended ranges to evaluate the behaviour of the proposed hashing framework under extreme image degradations. A total of 5440 distorted fingerprint samples were generated using stronger rotation, noise, blur, and scaling transformations than those used in the robustness experiments. The experimental results in Table 6 show that the normalized Hamming distance increases significantly under such severe distortions, with average values ranging from 0.53 to 0.55. These higher Hamming distances indicate that the generated hashes deviate substantially from the original fingerprint hashes when distortions exceed acceptable limits. Such behaviour is desirable for biometric hashing systems, as it ensures that heavily distorted or manipulated fingerprints are not incorrectly authenticated as genuine samples. Among the tested distortions, rotation produced the highest average Hamming distance, whereas noise resulted in the lowest values, indicating that the proposed hashing scheme retains partial tolerance to noise while remaining sensitive to strong geometric variations. These results further validate the effectiveness of the decision threshold used in the hashing framework, as the Hamming distances produced by extreme distortions remain well separated from the genuine matching distribution. Consequently, the proposed hashing scheme maintains reliable discrimination between legitimate fingerprint variations and heavily distorted or manipulated samples.Table 6Extreme distortion analysis.Distortion typeMean HDStd devMin HDMax HDNoise0.53410.03080.45250.6181Blur0.53860.03760.48330.8239Rotation0.54930.02560.47470.6229Scaling0.54460.03220.36000.7726 Table 6Extreme distortion analysis.Distortion typeMean HDStd devMin HDMax HDNoise0.53410.03080.45250.6181Blur0.53860.03760.48330.8239Rotation0.54930.02560.47470.6229Scaling0.54460.03220.36000.7726 Extreme distortion analysis. Discriminability analysis In order to measure the discriminability of the fingerprint hashing algorithm, FNMR, FMR, and EER are calculated for the sample fingerprint databases. Since each manipulation is tested against three different parameter values, the total number of images analyzed for each manipulation is 680 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} 3 = 2040. Figures 12 and 13 show performance trends observed in the experimental analysis to visualize how each manipulation affects performance. From the trend of variation of FNMR, FMR, and EER, it is evident that the values of these metrics are undesirable for other threshold values with respect to the data generated using the proposed algorithm. This also plays a role in the selection of the threshold value which balances the aforementioned metrics. Thus, these plots provide a better understanding of how each manipulation affects the algorithm’s capability to distinguish between genuine and imposter images, thus underlining the importance of optimal manipulation techniques in ensuring a secure and robust fingerprint recognition system.Fig. 12FNMR and FMR curves for image manipulations.Fig. 13EER curves for image manipulations. FNMR and FMR curves for image manipulations. EER curves for image manipulations. The results obtained from the fingerprint manipulation experiments demonstrate varying impacts on the discriminative capability and robustness of the proposed hashing algorithm, as summarized in Table 7. Among the evaluated manipulations, rotation and median filtering exhibit the most favorable performance, achieving low EER values (0.4236% and 0.3864%, respectively) along with minimal FNMR, indicating strong robustness while preserving discriminative fingerprint characteristics. Notably, rotation achieves a zero FNMR, highlighting the algorithm’s strong tolerance to rotational variations. Gamma correction and scaling also demonstrate acceptable robustness, with moderate EER values, indicating that the hashing framework can effectively accommodate photometric and geometric intensity variations without significant degradation in recognition performance. In contrast, JPEG compression shows the highest EER (2.9386%), suggesting a comparatively higher sensitivity to lossy compression artifacts, which may affect fine ridge structures and local texture information essential for accurate fingerprint matching. Overall, the observed balance between FMR and FNMR across different attack scenarios confirms the robustness of the proposed hashing scheme against common fingerprint manipulations. Minor variations between FNMR and FMR values can be observed more clearly upon zooming into the corresponding regions in the PDF, especially near the operating threshold, where subtle trade-offs between security and usability become evident.Table 7FNMR, FMR, EER, and corresponding EER thresholds for different fingerprint manipulations.ManipulationsFMR (%)FNMR (%)EER (%)EER thresholdGamma Correction2.53380.29410.5147724.7184Gaussian Filtering1.35850.14710.7528798.6273JPEG Compression3.08320.29412.9386750.3812Median Filtering0.62580.14710.3864798.6273Rotation0.84710.00000.4236798.6273Scaling0.86570.44120.6534798.6273 Table 7FNMR, FMR, EER, and corresponding EER thresholds for different fingerprint manipulations.ManipulationsFMR (%)FNMR (%)EER (%)EER thresholdGamma Correction2.53380.29410.5147724.7184Gaussian Filtering1.35850.14710.7528798.6273JPEG Compression3.08320.29412.9386750.3812Median Filtering0.62580.14710.3864798.6273Rotation0.84710.00000.4236798.6273Scaling0.86570.44120.6534798.6273 FNMR, FMR, EER, and corresponding EER thresholds for different fingerprint manipulations. Noise resilience Fingerprint images are often subjected to noises due to variations in sensor quality, environmental conditions, and acquisition imperfections. A robust fingerprint hashing algorithm should resist such distortions so slight variances in noise do not impact the generated hash. To analyze the noise resilience of the fingerprint hashing algorithm, salt-and-pepper and speckle noise are introduced into fingerprint images, and their impact is measured using Hamming distances. Salt-and-pepper noise introduces impulse noise via the random replacement of pixel values by black (0) or white (255). It can be created due to dust, scratches, or transmission errors that could arise during fingerprint acquisition. Speckle noise introduces granular distortions, mostly due to compression artifacts and sensor defects in fingerprint images. This noise is multiplicative, scaling pixel intensities rather than simply substituting one value for another. The noise density for Salt-and-pepper noise and the variance for Speckle noise are set to 2% and 0.01 respectively, assuring a suitably realistic level of noise without degradation to the image. After the addition of noises, hashes for the 680 fingerprint images are generated, and Hamming distances are calculated with respect to the original images. If the Hamming distance is within the 800-bit threshold, the algorithm successfully preserves fingerprint identity despite noise distortions. The average Hamming distance for salt-and-pepper noise is 526.4118 and that for speckle noise is 356.8323. The results, therefore, confirm that the fingerprint hashing algorithm is robust to natural noise distortions while preserving the identity of a fingerprint. This is crucial in most real-life applications, where noise cannot be avoided due to external factors. The impact of noise on Hamming distances can be seen in Fig. 14, and this further strengthens the algorithm’s performance and effectiveness.Fig. 14Noise resilience analysis. Noise resilience analysis. Comparison Manipulation tolerances To further evaluate the strength and discriminability of the proposed fingerprint hashing algorithm, a comparative analysis against existing studies that have explored fingerprint hashing and biometric feature extraction under numerous content-preserving manipulations is conducted. By evaluating the maximum manipulations reported in previous studies, the ability of the proposed algorithm to preserve essential fingerprint features while maintaining high discriminability between genuine and impostor fingerprints is examined. The Table 8 summarizes maximum manipulation parameter values applied to this study (see table 5) and compares them to those in relevant fingerprint hashing and biometric robustness studies.Table 8Comparison of previous studies based on various manipulations.ManipulationsParameterWang et al.15Muthu et al.16Abdullahi et al.20Birouk et al.23Proposed algorithmGamma correction\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}-1.3–0.75–1.50.75–1.30.9–1.05 (except 1)Gaussian filter\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}-0.5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}-0.850.001–0.0040.5–50.4–0.95JPEG compressionQuality\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}-10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}-1020–5010–9030–50Median filterFilter size–3x3–3x33x3–3x3–2x2–5x5Rotation\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}-10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{\circ }$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}-5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{\circ }$$\end{document}7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{\circ }$$\end{document}-20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{\circ }$$\end{document}1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{\circ }$$\end{document}-5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{\circ }$$\end{document}15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{\circ }$$\end{document}-25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{\circ }$$\end{document}ScalingScale factor\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}-0.5–0.5–2–0.8–1.05 (except 1)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document} denotes that a single parameter value is considered in the article. Table 8Comparison of previous studies based on various manipulations.ManipulationsParameterWang et al.15Muthu et al.16Abdullahi et al.20Birouk et al.23Proposed algorithmGamma correction\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}-1.3–0.75–1.50.75–1.30.9–1.05 (except 1)Gaussian filter\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}-0.5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}-0.850.001–0.0040.5–50.4–0.95JPEG compressionQuality\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}-10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}-1020–5010–9030–50Median filterFilter size–3x3–3x33x3–3x3–2x2–5x5Rotation\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}-10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{\circ }$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}-5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{\circ }$$\end{document}7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{\circ }$$\end{document}-20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{\circ }$$\end{document}1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{\circ }$$\end{document}-5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{\circ }$$\end{document}15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{\circ }$$\end{document}-25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{\circ }$$\end{document}ScalingScale factor\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document}-0.5–0.5–2–0.8–1.05 (except 1)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document} denotes that a single parameter value is considered in the article. Comparison of previous studies based on various manipulations. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim$$\end{document} denotes that a single parameter value is considered in the article. The proposed algorithm uses a controlled way of optimizing the content with robustness against different image manipulations. This is subsequently more consistent with large rotation variability and using a wider Gaussian filtering range to avoid large values that could distort image characteristics. Unlike studies that explore a wide spectrum of JPEG compression, the proposed method focuses on moderate levels (30-50), enhancing integrity. Similarly, proper selection of gamma value and scaling factor prevents excessive distortions that also allow for further adaptability. The broadened median filtering range allows for higher robustness against noise while still at least preserving the details of the images. Such comparative assessment becomes essential since biometric systems must be resistant to real-world distortions while still maintaining secure, consistent, and distinguishable representations of fingerprint data. If an algorithm is sensitive to minor modifications, then it may produce inconsistent hashes, leading to higher false rejection rates. On the contrary, if it is not sufficiently discriminated, it may not be able to discriminate between distinct fingerprints, thus yielding increased false acceptance rates. Through benchmarking against prior studies, the proposed fingerprint hashing method demonstrates strong resistance to extensive transformations while preserving the uniqueness of each fingerprint hash, indicating its suitability for robust biometric template protection applications. Feature preservation and image quality metrics It is extremely important that the system retains the key aspects of a fingerprint while reducing noise. All transformations applied for enhancement should not change the structural integrity, edge details, and minutiae features to ensure reliable recognition. The performance analysis of this approach employing FWT is compared with Gaussian filtering in terms of noise reduction, feature retention, and edge preservation by considering some key image quality metrics such as Signal-to-Noise Ratio (SNR)37, Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index (SSIM). Gaussian filtering works meticulously in the removal of Gaussian noise by averaging the pixel intensities using a Gaussian kernel. Due to blurring, it tends to smooth out the fine ridge structures, and loss of this detail in the minutiae reduces fingerprint recognition, as edges and local textures are imperative for the same. Unlike a Gaussian filter, which works in the spatial domain, FWT works in the frequency domain, and it can therefore remove noise while preserving sharp edges and fine detail. This makes FWT particularly effective for fingerprint images, where maintaining ridge structures is crucial. To point this out,the Edge Preservation Index is also calculated for both techniques. In addition to the above metrics, entropy and mutual information (MI)39 for FWT- and Gaussian-filtered images are also computed to further evaluate the feature preservation and information content. Entropy measures the overall uncertainty or complexity in the fingerprint image: higher entropy indicates richer texture and preserved structural details, which are crucial for reliable recognition. Mutual Information quantifies the statistical dependency between the original and denoised images; higher MI indicates better retention of essential fingerprint features while removing noise. The experimental analysis demonstrates that FWT outperforms Gaussian filtering in terms of preserving the fingerprint details and noise reduction. Figure 15 presents the original fingerprint image, its denoised versions using FWT and Gaussian filtering, and their corresponding edge maps. Comparative results (Table 9) show the consistent superior performance of FWT across all metrics, including Entropy and MI. PSNR and SNR of FWT-processed images confirm that FWT minimizes distortion while maintaining higher fidelity concerning the structure of fingerprints than Gaussian-filtered images. The high and consistent SSIM once again shows FWT’s feasibility and power in retaining structural features in fingerprint images. The higher entropy and MI values for FWT further indicate that the denoising process retains richer information content and stronger correlation with the original fingerprint image than Gaussian filtering, which is vital for subsequent recognition tasks.Fig. 15Original fingerprint, denoised (FWT and Gaussian), and edge maps. Original fingerprint, denoised (FWT and Gaussian), and edge maps. Gaussian filtering, besides being a popular denoising approach, also smoothens the image, which tends to produce a lack of ridge sharpness and degradation of important features. The lower values for Edge Preservation Index, SSIM, Entropy, and MI suggest that Gaussian filtering sacrifices critical fingerprint details and reduces information content. The graphical representations in Fig. 16 justify these findings, indicating that throughout all tests, FWT has consistently surpassed Gaussian filtering.Table 9Average metric values for FWT and Gaussian filtering.MetricsFWTGaussian filteringPSNR70.629863.9295SNR19.09712.5588SSIM0.94080.7107Edge Preservation Index0.86570.5933Entropy6.73076.4731Mutual information3.12172.8258 Table 9Average metric values for FWT and Gaussian filtering.MetricsFWTGaussian filteringPSNR70.629863.9295SNR19.09712.5588SSIM0.94080.7107Edge Preservation Index0.86570.5933Entropy6.73076.4731Mutual information3.12172.8258 Average metric values for FWT and Gaussian filtering. Fig. 16Graphical representation of metrics. Graphical representation of metrics. Computational efficiency and performance analysis To evaluate the computational efficiency of the proposed framework, a comprehensive execution time analysis was performed and compared with the existing algorithm presented in29. To ensure a fair benchmarking procedure, the existing algorithm was reimplemented in the same MATLAB environment and executed on the identical hardware platform used for the proposed method. Both implementations followed comparable preprocessing configurations and dataset inputs in order to maintain consistent experimental conditions. From these experiments, it can be observed that the proposed algorithm requires significantly less execution time than the existing approach. The average implementation time of the proposed method is approximately 2.507 s, whereas the existing algorithm requires approximately 1048.888 s. This substantial reduction in execution time primarily results from the use of the Fast Wavelet Transform (FWT) within the fractal compression stage, which effectively reduces the computational complexity associated with feature extraction and representation. The reduction in computational time significantly improves the practicality of the proposed fingerprint hashing framework for handling larger biometric datasets. Faster processing enables efficient fingerprint analysis and reduces computational bottlenecks that may arise in large biometric repositories. Figure 17 illustrates the comparison of execution times between the proposed algorithm and the existing algorithm, highlighting the considerable computational advantage of the proposed approach.Fig. 17Execution time comparison. Execution time comparison. Beyond computational efficiency, the False Match Rate (FMR) and False Non-Match Rate (FNMR) of the proposed fingerprint hashing algorithm are also compared with the existing algorithm. This comparison aims to evaluate whether the proposed approach maintains reliable discriminability while improving computational efficiency. The FNMR and FMR values of the proposed algorithm are provided in Table 7, while the corresponding values for the existing algorithm are shown in Table 2. Figure 18 presents a schematic comparison of the FNMR and FMR values for both methods.Fig. 18Comparison of FNMR and FMR for proposed and existing algorithms. Comparison of FNMR and FMR for proposed and existing algorithms. For statistical validation, the mean and standard deviation of FMR and FNMR were computed across all manipulations and datasets for both the proposed and existing methods. Paired statistical tests, including a t-test and a Wilcoxon signed-rank test, were performed to evaluate whether the observed differences are statistically significant. As shown in Table 10, the differences between the proposed and existing methods are not statistically significant at the 0.05 significance level. This indicates that the proposed algorithm achieves performance comparable to the existing approach in terms of FMR and FNMR.Table 10Statistical validation of FMR and FNMR for the proposed and existing methods.MetricExisting (mean ± std)Proposed (mean ± std)t-test p-valueWilcoxon p-valueFMR (%)0.93 ± 1.111.55 ± 1.020.41120.4375FNMR (%)1.53 ± 2.190.22 ± 0.150.20960.3125 Table 10Statistical validation of FMR and FNMR for the proposed and existing methods.MetricExisting (mean ± std)Proposed (mean ± std)t-test p-valueWilcoxon p-valueFMR (%)0.93 ± 1.111.55 ± 1.020.41120.4375FNMR (%)1.53 ± 2.190.22 ± 0.150.20960.3125 Statistical validation of FMR and FNMR for the proposed and existing methods. The discriminability metrics (FMR and FNMR) of the proposed method therefore remain broadly comparable to those reported for the existing algorithm. Consequently, the primary empirical advantage of the proposed framework lies in its substantial improvement in computational efficiency while maintaining competitive recognition performance. Furthermore, the proposed algorithm was evaluated across multiple fingerprint datasets, including FVC2002, FVC2004, and SOCOFing. In particular, the proposed method maintains low FNMR values across various manipulations such as median filtering, rotation, and scaling, while the FMR remains controlled across most attack scenarios, with only a slight increase observed under JPEG compression. These observations indicate that the proposed framework provides stable and computationally efficient fingerprint hashing performance across diverse datasets and manipulation conditions. Comparison with learning-based fingerprint hashing methods To address concerns regarding the absence of learning-based baselines, the proposed fingerprint hashing framework is compared with representative deep learning–based fingerprint recognition and template protection approaches reported in recent literature. While most learning-based methods are primarily designed to maximize recognition accuracy and typically report Equal Error Rate (EER) under standard evaluation settings, the objective of the present comparison is to provide a broader methodological perspective rather than strict performance equivalence. It should be noted that the results for learning-based approaches are obtained from their respective published studies and were evaluated using different datasets and experimental protocols. Therefore, the comparison presented here is intended to provide contextual insight into the characteristics of different fingerprint representation approaches and should be interpreted qualitatively rather than as a controlled empirical benchmark. In addition to EER, key practical aspects such as training requirements, computational runtime characteristics, and template representation size are also considered. Table 11 summarizes these factors for several representative learning-based fingerprint models alongside the proposed transform-based hashing framework, highlighting the trade-offs between accuracy-oriented learning approaches and lightweight transform-based template protection methods.Table 11Comparison of the proposed method with representative learning-based fingerprint approaches.CriteriaPoreNet26Siamese CNN + SIFT5Cancelable CNN27ProposedTypeCNN pore descriptorSiamese CNNDeep cancelable biometricTransform-based hashingTraining requiredYesYesYesNoRuntimeHighHighHighVery LowHash / template lengthFeature vectorLearned embeddingBinary template4096-bit hashDataset(s)PolyU HRF DBI & DBIIFVC2002CMBD, CASIA Iris V3 & FVC2002 DB2FVC2002/04 & SOCOFingEER (%)2.91 (partial) / 0.57 (complete)3.900.038–0.0730.8862 Table 11Comparison of the proposed method with representative learning-based fingerprint approaches.CriteriaPoreNet26Siamese CNN + SIFT5Cancelable CNN27ProposedTypeCNN pore descriptorSiamese CNNDeep cancelable biometricTransform-based hashingTraining requiredYesYesYesNoRuntimeHighHighHighVery LowHash / template lengthFeature vectorLearned embeddingBinary template4096-bit hashDataset(s)PolyU HRF DBI & DBIIFVC2002CMBD, CASIA Iris V3 & FVC2002 DB2FVC2002/04 & SOCOFingEER (%)2.91 (partial) / 0.57 (complete)3.900.038–0.0730.8862 Comparison of the proposed method with representative learning-based fingerprint approaches. In contrast to learning-based fingerprint recognition frameworks, the proposed method does not require model training or large-scale labelled datasets. Most deep learning approaches rely on computationally intensive convolutional architectures and extensive training procedures to achieve high discriminative performance. Although such models can obtain very low Equal Error Rates under controlled evaluation conditions, they typically generate high-dimensional feature embeddings and involve significant runtime and memory requirements. The proposed framework, in contrast, employs a deterministic transform-based pipeline that combines FWT, FMT, and fractal coding to generate a fixed-length 4096-bit fingerprint hash without requiring a training stage. This design leads to reduced computational complexity and faster execution compared with typical learning-based approaches, while maintaining reasonable recognition performance across the evaluated benchmark datasets. Consequently, the proposed method is well-suited for biometric template protection scenarios, where properties such as fixed-length hashing, revocability, and computational efficiency are important design considerations. Computational complexity analysis The computational cost of the proposed framework is primarily dominated by the fractal encoding stage, which requires matching each range block with candidate domain blocks. In a conventional fractal coding scheme, each range block is compared with all domain blocks, resulting in a computational complexity of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(R_n \times D_n)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R_n$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D_n$$\end{document} denote the number of range and domain blocks, respectively. In the present implementation, the fingerprint image is partitioned into 64 range blocks, while the domain pool contains 2401 domain blocks, leading to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$64 \times 2401 = 153{,}664$$\end{document} matching comparisons during the encoding stage. To reduce this computational burden, the proposed framework incorporates Fast Wavelet Transform (FWT)-based filtering prior to fractal matching. Domain blocks whose wavelet energy characteristics differ significantly from the range block are discarded, thereby restricting the search to a small set of candidates. In practice, the candidate set reduces to approximately 10–15 domain blocks, resulting in only \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$64 \times 15 = 960$$\end{document} comparisons. This wavelet-guided candidate selection reduces the fractal search space from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D_n$$\end{document} to a significantly smaller subset K, leading to an optimized matching complexity of18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} O(R_n K), \quad K \ll D_n \end{aligned}$$\end{document}Consequently, the proposed framework transforms the exhaustive fractal search into a guided matching process, significantly lowering computational cost while preserving discriminative fingerprint features. The complexity of the major stages of the algorithm is summarized in Table 12.Table 12Computational complexity of the major stages of the proposed fingerprint hashing framework.StageComputational complexityImage preprocessing\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(N^2)$$\end{document}Fast wavelet transform (FWT)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(N^2)$$\end{document}Fourier–Mellin transform (FMT)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(N^2 \log N)$$\end{document}Optimized fractal matching\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(R_n K),\; K \ll D_n$$\end{document} Table 12Computational complexity of the major stages of the proposed fingerprint hashing framework.StageComputational complexityImage preprocessing\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(N^2)$$\end{document}Fast wavelet transform (FWT)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(N^2)$$\end{document}Fourier–Mellin transform (FMT)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(N^2 \log N)$$\end{document}Optimized fractal matching\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(R_n K),\; K \ll D_n$$\end{document} Computational complexity of the major stages of the proposed fingerprint hashing framework. Security and privacy analysis Building on the robustness, discriminability, and efficiency results reported in “Efficiency analysis”, this section provides a dedicated analysis of the security and privacy properties of the proposed fingerprint hashing framework under a standard biometric template protection threat model. Threat model and security metrics The security of a biometric hashing system depends on its ability to protect stored templates against unauthorized access and to resist active attacks during authentication. In this work, a commonly adopted threat model for biometric template protection is considered. It is assumed that an adversary may gain access to the stored biometric hash templates and may observe multiple authentication attempts, but does not possess knowledge of the internal system parameters such as wavelet decomposition configurations, threshold values, or fractal encoding parameters. Under this threat model, several potential attack scenarios are analyzed using quantitative security metrics in order to evaluate the robustness of the proposed hashing framework. Template leakage and inversion attacks Template leakage and inversion attacks attempt to reconstruct the original biometric image or infer discriminative biometric information from stored hash values. Since the proposed framework applies multiple irreversible feature transformations followed by fixed-length hashing, direct reconstruction of the original fingerprint image is highly difficult in practice. The irreversibility of the proposed representation is primarily achieved through a sequence of transformations during feature extraction. The fingerprint image is first processed using the Fast Wavelet Transform (FWT), which captures multi-resolution texture information. This is followed by the Fourier–Mellin Transform (FMT) to achieve rotation and scale normalization. Subsequently, the fractal encoding stage performs block-based domain–range mapping and quantization, which compresses structural information while discarding precise spatial relationships. These sequential transformations eliminate direct spatial correspondence between the original fingerprint image and the final hash representation, thereby making reconstruction of the original biometric sample from the stored hash computationally challenging without access to intermediate representations. To quantitatively evaluate resistance against such attacks, two statistical measures are analyzed: hash entropy and mutual information between the original fingerprint images and the generated hash codes. High entropy indicates that the hash values are well distributed and unpredictable, while low mutual information suggests limited information leakage from the biometric data to the stored template. The entropy of the generated hash values is evaluated using normalized Shannon entropy. For a binary hash sequence H, the entropy is computed as\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E = - \sum _{i=0}^{1} p_i \log _2(p_i)$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_0$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_1$$\end{document} denote the probabilities of occurrence of bit values 0 and 1 in the generated hash sequence. Since the entropy is calculated for binary symbols, the maximum possible entropy is 1 bit per hash bit, which occurs when both bit values are equally probable. In the proposed framework, the biometric template is represented as a fixed-length binary hash of 4096 bits. Under the assumption of an ideally uniform distribution, the theoretical maximum entropy of the hash space is therefore approximately 4096 bits, providing a large search space against brute-force guessing or template enumeration attacks. The experimentally obtained entropy value of 0.8931 represents the average normalized per-bit Shannon entropy computed from the generated 4096-bit hashes across the 680 fingerprint samples used in the experimental evaluation. A value approaching unity indicates that the hash bits are well distributed and contain high information content, which reduces predictability and strengthens resistance against template inversion attempts. In addition, the mean mutual information value of 0.5673 indicates a moderate statistical dependency between the fingerprint data and the generated hash representation, suggesting that only partial information from the original biometric signal is retained in the final hash codes. Together, these results suggest that the proposed hashing scheme reduces the amount of biometric information exposed through stored templates, thereby improving resistance to inversion and reconstruction attacks. Replay attacks Replay attacks involve reusing previously captured biometric data or hash values in an attempt to impersonate a legitimate user. In a secure biometric hashing system, variations in acquisition conditions or input perturbations should produce sufficiently different hash outputs to prevent successful replay. To evaluate replay attack resistance, the Hamming distance between hash codes generated from replayed or perturbed inputs is computed. The experimental evaluation yielded a mean replay Hamming distance of 408.05, which is significantly higher than typical matching thresholds used in biometric systems. This indicates that replayed biometric samples are unlikely to generate matching hash codes, thereby reducing the probability of successful replay-based impersonation. Cross-matching attacks Cross-matching attacks attempt to link biometric templates across different databases or applications in order to track user identities. An effective template protection scheme should ensure that hash codes generated from different fingerprints are statistically independent and sufficiently separated. To assess resistance against cross-matching, inter-class Hamming distance analysis is conducted. The results show a mean inter-class Hamming distance of 752.50, indicating a high degree of separation between hash codes corresponding to different fingerprint identities. Such large inter-class distances reduce the likelihood of accidental matches and help support unlinkability across independent databases. The combined analysis of entropy, replay resistance, mutual information, and inter-class separability indicates that the proposed hashing framework provides a reasonable level of protection against template leakage, replay, inversion, and cross-matching attacks. Table 13 and Fig. 19 illustrates the comparative performance of the proposed framework across multiple evaluation metrics including PSNR, SNR, SSIM, entropy, and mutual information. The results demonstrate that the proposed FWT-based framework consistently preserves fingerprint structural information while effectively suppressing noise. High entropy and large Hamming distance values indicate strong unpredictability and separability of the generated hashes, while relatively low mutual information suggests limited biometric information leakage from the original fingerprint images to the stored hash templates.Table 13Average metric values for threat and security analysis.MetricsMean valueHash entropy0.8931Replay hamming distance408.05Mutual information0.5673Mean inter-class HD752.50 Table 13Average metric values for threat and security analysis.MetricsMean valueHash entropy0.8931Replay hamming distance408.05Mutual information0.5673Mean inter-class HD752.50 Average metric values for threat and security analysis. Fig. 19Graphical representation of metric values. Graphical representation of metric values. Collision probability analysis To evaluate the distinctiveness of the generated hash codes, an inter-class collision analysis is conducted by computing pairwise Hamming distances between hash codes belonging to different fingerprint identities. A collision is defined as a pair of hashes whose Hamming distance falls below the predefined matching threshold. From a theoretical perspective, the collision probability of an ideal binary hash function depends on the length of the hash representation. For a uniformly distributed 4096-bit hash space, the probability that two randomly generated hashes collide can be approximated as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{collision}=2^{-4096}$$\end{document}. Although practical biometric hashes are not perfectly random due to inherent biometric correlations, this theoretical bound highlights the extremely large size of the available hash space and indicates that accidental collisions are highly unlikely under ideal random hash assumptions. Based on the experimental evaluation, 5,462 collisions are observed out of 229,503 inter-class comparisons, resulting in an empirical collision rate of 0.0238 (2.38%). These comparisons were computed using the fingerprint samples selected from the SOCOFing and FVC datasets used in the experimental evaluation. This result indicates a measurable separation between intra-class similarity and inter-class dissimilarity in the generated hash space. In biometric authentication systems, template comparisons are typically performed using threshold-based similarity measures such as the False Match Rate (FMR), where limited near-collision behaviour may arise depending on the system parameters and database size40. Consequently, small inter-class overlaps can occur in practical template spaces while still maintaining reliable recognition performance. Nevertheless, further reduction of collision probability remains an important direction for future research. It is important to note that some implementations report non-collision probability (1-collision rate). However, in this work, results are explicitly reported in terms of empirical collision rate for clarity and reproducibility. While deep learning–based template protection schemes may offer lower collision rates under large-scale training, the proposed handcrafted approach provides a favourable balance between security, interpretability, and computational efficiency, making it suitable for practical biometric template protection applications. Revocability and unlinkability analysis Revocability and unlinkability are essential requirements for biometric template protection. Revocability ensures that compromised templates can be replaced, while unlinkability prevents cross-matching of templates generated from the same biometric under different system parameters. These properties are essential requirements in modern biometric template protection standards and ensure that biometric identifiers can be securely managed even if stored templates are exposed. In the proposed framework, revocability is achieved by modifying system parameters such as wavelet decomposition levels, threshold values, and fractal encoding parameters, which produce significantly different hash outputs for the same fingerprint. To quantify unlinkability, the Hamming distance between hash codes generated from the same fingerprint under different parameter settings is computed. The experimental results show a mean unlinkability Hamming distance of 728.1043, indicating a high degree of dissimilarity between regenerated hash codes. This confirms that the proposed hashing scheme effectively prevents cross-matching and supports template revocation without compromising recognition performance. Conclusion and future scope This work presents a structured and computationally efficient fingerprint hashing framework that integrates established transform techniques such as Fast Wavelet Transform (FWT), Fourier–Mellin Transform (FMT), and fractal coding to enhance noise resilience while maintaining high efficiency. The FWT component plays a crucial role by providing superior noise suppression compared to conventional Gaussian filtering, effectively handling diverse noise patterns through multiresolution decomposition while preserving discriminative fingerprint features. Additionally, FWT accelerates the fractal compression stage by enabling efficient domain–range block matching in the wavelet domain, significantly reducing computational overhead without compromising hashing accuracy. Extensive experiments conducted on publicly available datasets, including FVC2002 fingerprint database, FVC2004 fingerprint database, and SOCOFing fingerprint dataset, demonstrate that the proposed method produces a compact 1049-bit hash while achieving robustness to noise and geometric transformations, making it suitable for practical fingerprint-based security applications. From a broader perspective, the framework offers a lightweight alternative to learning-based methods such as PoreNet fingerprint model, Siamese CNN fingerprint architecture, and Cancelable CNN fingerprint templates by eliminating the need for large-scale training, complex optimization, and high computational resources, while still maintaining competitive discriminability; however, such comparisons are contextual rather than empirical due to differences in datasets and protocols. Overall, the proposed approach emphasizes computational efficiency, robustness, and explainability, and future work will focus on extending resilience to more challenging noise models such as impulsive and non-Gaussian noise, as well as exploring its applicability to other biometric modalities including iris, face, and voice, thereby enabling the development of unified multimodal authentication systems that balance efficiency, security, and robustness.