Transformer-based iris verification with attention-guided segmentation and Siamese learning Authors: S. Ramesh, V. Krishnaveni Published: 2026 ABSTRACT This study presents a transformer-based Siamese framework for iris verification that integrates image enhancement, attention-guided segmentation, and global feature learning within a unified pipeline. A deep denoising autoencoder is employed to improve image quality, followed by an attention-guided U-Net for precise iris localization. Global feature representations are extracted using a Vision Transformer, enabling the modeling of long-range dependencies across iris textures, and are optimized through contrastive learning to enhance discriminability. The proposed framework is evaluated on the CASIA-IrisV3 dataset, including the Interval, Lamp, and Twins subsets, under a subject-disjoint verification protocol. Experimental results achieve an Equal Error Rate (EER) of 2.34% and an Area Under the ROC Curve (AUC) of 0.987, with a True Acceptance Rate of approximately 95% at a False Acceptance Rate of 10⁻³. The method attains a verification accuracy of 97.82% while maintaining efficient inference with an average latency of 12.6 ms per comparison. These results indicate consistent performance across varying acquisition conditions within the near-infrared domain, demonstrating the effectiveness of integrating attention-guided segmentation and transformer-based representations for iris verification. Introduction Biometric authentication has become a fundamental component of modern identity management systems due to its ability to provide intrinsic and non-transferable identity cues. Common modalities, including fingerprint, face, voice, and iris recognition, each offer distinct advantages and operational trade-offs1. Recent advances have also explored neural signals such as electroencephalography (EEG) for identity recognition2. Among these modalities, iris recognition is widely regarded as one of the most reliable due to its high inter-subject variability, structural richness, and long-term stability1,3. The complex micro-patterns within the iris—such as crypts, furrows, radial streaks, and collarette structures—form a highly distinctive feature space that supports fine-grained discrimination4. Compared to facial appearance, iris texture is less affected by aging and expression variation, making it particularly suitable for high-assurance verification scenarios3. Despite these advantages, reliable iris verification remains challenging under practical acquisition conditions. Variations in illumination, occlusion by eyelids or eyelashes, motion blur, and pupil dilation can introduce structural distortions that degrade representation stability5,6. Early iris recognition systems relied on handcrafted encoding schemes combined with precise normalization; however, their sensitivity to segmentation inaccuracies and imaging noise limited robustness1,7. Deep convolutional neural networks (CNNs) have improved adaptive feature extraction8–10, yet their localized receptive fields may limit the modeling of long-range structural dependencies across the iris annulus11. Transformer architectures, driven by self-attention mechanisms, provide a principled approach for modeling global contextual relationships by dynamically capturing interactions across spatial regions. However, directly applying global encoders to imperfectly localized iris regions may degrade embedding quality, as structural inconsistencies, background interference, and periocular artifacts can propagate through the network and reduce discriminative separability between genuine and impostor comparisons12. These observations suggest that enhancement, spatial isolation, and embedding learning should be addressed jointly within a unified framework. It is important to distinguish the roles of attention within the proposed approach. Spatial attention is incorporated in the segmentation stage through an attention-gated U-Net to improve region-of-interest extraction. In contrast, the Vision Transformer employs standard self-attention for global feature modeling without explicit modulation. The proposed framework leverages these complementary mechanisms to improve both spatial localization and contextual representation. Motivated by these considerations, we propose an integrated iris verification framework that combines reconstruction-based enhancement, attention-guided segmentation, and transformer-based embedding learning within a metric-learning paradigm. A deep denoising autoencoder is used to improve structural consistency prior to localization. An attention-gated U-Net extracts the iris region to suppress peripheral interference. A Vision Transformer encodes global contextual dependencies, and representation geometry is optimized through Siamese contrastive learning to enforce intra-class compactness and inter-class separation. The framework is evaluated using a strictly subject-disjoint protocol on the CASIA-IrisV3 dataset (Interval, Lamp, and Twins subsets) to ensure unbiased verification assessment. Experimental results demonstrate strong separation between genuine and impostor similarity distributions, improved verification performance, and computational efficiency suitable for near-infrared biometric systems under controlled and semi-controlled conditions. Unlike prior approaches that apply transformer-based embedding in isolation, the proposed framework integrates enhancement and attention-guided segmentation prior to global feature learning, reducing the propagation of structural noise into the embedding space. The main contributions of this work are: A unified verification architecture integrating enhancement, attention-guided segmentation, and transformer-based embedding learning.Statistical analysis of embedding separability under a subject-disjoint evaluation protocol.Comprehensive ablation study quantifying the individual and combined contributions of key components.Empirical evaluation demonstrating a balance between verification performance and computational efficiency. A unified verification architecture integrating enhancement, attention-guided segmentation, and transformer-based embedding learning. A unified verification architecture integrating enhancement, attention-guided segmentation, and transformer-based embedding learning. Statistical analysis of embedding separability under a subject-disjoint evaluation protocol. Statistical analysis of embedding separability under a subject-disjoint evaluation protocol. Comprehensive ablation study quantifying the individual and combined contributions of key components. Comprehensive ablation study quantifying the individual and combined contributions of key components. Empirical evaluation demonstrating a balance between verification performance and computational efficiency. Empirical evaluation demonstrating a balance between verification performance and computational efficiency. Related work Recent advances in iris recognition span multiple directions, including handcrafted feature learning, deep convolutional models, segmentation-oriented frameworks, transformer-based representations, and multimodal or security-focused systems. For clarity, prior work is organized into these categories. Early studies primarily focused on handcrafted feature extraction and classical learning approaches. Stylometric feature representations combined with machine learning were investigated by Adamovic et al.13, while Ahmadi et al.14 explored supervised classification strategies for biometric recognition. Radial basis function (RBF) networks were also applied for iris discrimination15. Beyond iris recognition, broader biometric modalities such as EEG-based authentication have been explored for secure identity verification2. Although these approaches established foundational feature engineering techniques, they are generally limited in capturing the complex and highly variable texture patterns of the iris. With the emergence of deep learning, convolutional neural networks (CNNs) became the dominant paradigm for feature extraction. Abdellatef et al.16 proposed a deep learning-based cancelable face–iris system, while Alaslni and Elrefaei8 demonstrated transfer learning strategies using CNNs. Multimodal fusion approaches combining face and iris modalities were explored to improve robustness17. CNN-based recognition was further studied by Azam and Rana9, and longitudinal performance challenges were analyzed by Das et al.3. Additional works focused on improving verification performance and computational efficiency through optimized architectures18,19. Deep embedding-based methods such as DeepIris20 and complex-valued networks11 enhanced representational richness, while feature optimization and multimodal descriptors were explored in7,21. CNN architectures, including VGG-based models22, further demonstrated strong discriminative capability. However, CNN-based approaches primarily focus on local feature extraction and may not explicitly capture long-range spatial dependencies across the iris structure. Accurate iris localization and image enhancement are critical for reliable recognition. Huo et al.12 proposed a U-Net++–based segmentation framework for precise boundary detection, while Jamaludin et al.5 addressed image degradation through deblurring techniques. Robust localization methods were introduced by Jan et al.6, and the impact of contact lenses on recognition was studied in23. High-throughput recognition using near-infrared sequences was explored in24. Additional improvements include CNN-based super-resolution25 and secure encrypted recognition frameworks for IoT environments26. While these approaches improve preprocessing quality, they typically treat enhancement and feature learning as separate stages. More recently, advanced architectures have been explored for richer feature representation. Capsule network-based iris recognition has been investigated to improve feature encoding, including approaches by Zhao et al.27 and Kuhifayegh and Rajabi28, which incorporate enhanced routing mechanisms for improved discrimination. Transformer-based approaches, such as IrisFormer29, leverage self-attention mechanisms to model global dependencies within iris textures. These methods address limitations of convolutional architectures by enabling global contextual modeling; however, they are often applied without tightly integrated preprocessing and localization stages. Security and robustness remain important considerations in biometric systems. Cancelable iris recognition using comb filtering was proposed in30, while hybrid biometric authentication for cloud environments was explored in31. Cross-spectral recognition32, parallel localization with deep verification33, and comprehensive system reviews1 highlight ongoing challenges in generalization and deployment. Lightweight CNN strategies have also been proposed to improve computational efficiency10. Recent developments in vision–language models, such as CLIP, have demonstrated strong generalization capabilities across visual recognition tasks by leveraging language-guided semantic priors. Several studies have applied CLIP-based frameworks to face anti-spoofing and domain generalization, improving robustness under cross-domain variations and unseen attack scenarios. Approaches such as FM-CLIP34 and FA³-CLIP35 integrate multimodal feature alignment and frequency-aware representations, while prompt learning strategies have been explored to enhance generalization36–39. Additional works incorporating semantic anchors39, unified datasets40, and vision–language semantic awareness41 further highlight the potential of semantic-guided representations. Although the present work focuses on iris verification, these studies indicate promising directions for improving generalization in biometric systems. Despite significant progress, several limitations remain. Many existing approaches treat enhancement, localization, and feature extraction as independent stages, which may propagate structural inconsistencies into similarity computation. Furthermore, while CNN-based models excel at local feature learning, explicit modeling of global spatial dependencies remains limited. These limitations motivate the development of unified frameworks that integrate structural normalization, attention-guided segmentation, and globally contextualized feature learning within a cohesive metric-learning paradigm. Methodology This study proposes a unified deep learning framework for iris verification that integrates preprocessing, segmentation, feature extraction, and metric-based pattern matching into a cohesive pipeline. The overall architecture of the proposed iris verification framework is illustrated in Fig. 1. Fig. 1Architecture of the Proposed Iris Recognition Framework. Architecture of the Proposed Iris Recognition Framework. The system is designed to enhance discriminative representation learning while maintaining computational efficiency suitable for real-time verification. A deep denoising autoencoder (DDAE) is first applied to improve image quality and reduce acquisition-induced noise. Precise iris localization is then performed using an attention-enhanced U-Net to isolate the region of interest. Discriminative feature embeddings are subsequently extracted using a pretrained Vision Transformer (ViT), enabling global contextual modeling of iris textures. Finally, a Siamese network performs similarity learning for verification, where identity decisions are determined based on a threshold derived from Receiver Operating Characteristic (ROC) analysis. The framework is evaluated on the CASIA-IrisV3 dataset (Interval, Lamp, and Twins subsets) under a subject-disjoint verification protocol to ensure unbiased performance assessment across identities. Problem formulation Iris recognition in this study is formulated as a binary verification problem, where the objective is to determine whether two iris samples correspond to the same identity. Let an iris image be represented as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:I\in\:{\mathbb{R}}^{H\times\:W}$$\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}$$\:H$$\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}$$\:W$$\end{document}denote spatial dimensions. The proposed framework aims to learn a nonlinear embedding function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{\theta\:}(\cdot\:)$$\end{document}, parameterized by network weights \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}, that maps each input image into a compact and discriminative latent representation:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:z={f}_{\theta\:}\left(I\right),z\in\:{\mathbb{R}}^{d}&\:&\:\end{array}$$\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}$$\:d$$\end{document} denotes the embedding dimensionality. The objective of the learning process is to minimize intra-class variation while maximizing inter-class separability within the embedding space. Given two iris samples \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{I}_{i}$$\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}$$\:{I}_{j}$$\end{document}, their similarity is computed using cosine similarity:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:S({z}_{i},{z}_{j})=\frac{{z}_{i}^{{\top\:}}{z}_{j}}{\parallel\:{z}_{i}{\parallel\:}_{2}\parallel\:{z}_{j}{\parallel\:}_{2}}&\:&\:\end{array}$$\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}$$\:{z}_{i}$$\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}$$\:{z}_{j}$$\end{document} are the corresponding embeddings. The similarity score lies in the range \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left[-\mathrm{1,1}\right]$$\end{document}, where higher values indicate stronger correspondence. The verification decision \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{D}_{ij}$$\end{document} is defined as:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{D}_{ij}=\left\{\begin{array}{cc}1,&\:\mathrm{if\:}S({z}_{i},{z}_{j})\ge\:\tau\:\\\:0,&\:\mathrm{otherwise}\end{array}\right.&\:&\:\end{array}$$\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}$$\:\tau\:$$\end{document} is the decision threshold determined from validation-based ROC analysis. Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{D}_{ij}=1$$\end{document} represents a genuine pair, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{D}_{ij}=0$$\end{document} denotes an impostor pair. During training, discriminative representation learning is achieved using contrastive loss:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:L={y}_{ij}\cdot\:{d}_{ij}^{2}+(1-{y}_{ij})\cdot\:\mathrm{m}\mathrm{a}\mathrm{x}(0,m-{d}_{ij}{)}^{2}&\:&\:\end{array}$$\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}$$\:{y}_{ij}\in\:\left\{\mathrm{0,1}\right\}$$\end{document} is the ground-truth label, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{ij}=\parallel\:{z}_{i}-{z}_{j}{\parallel\:}_{2}$$\end{document}is the Euclidean distance between embeddings, 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} is a predefined margin enforcing inter-class separation. This formulation encourages compact clustering of embeddings belonging to the same identity while penalizing embeddings of different identities that fall within the margin. Verification performance is evaluated using False Acceptance Rate (FAR) and False Rejection Rate (FRR):5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:\mathrm{FAR}=\frac{\mathrm{False\:Acceptances}}{\mathrm{Total\:Impostor\:Comparisons}}&\:&\:\end{array}$$\end{document}6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:\mathrm{FRR}=\frac{\mathrm{False\:Rejections}}{\mathrm{Total\:Genuine\:Comparisons}}&\:&\:\end{array}$$\end{document} The Equal Error Rate (EER) corresponds to the operating point where:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:\mathrm{FAR}\left(\tau\:\right)=\mathrm{FRR}\left(\tau\:\right)&\:&\:\end{array}$$\end{document} EER serves as a threshold-independent indicator of verification performance and is widely adopted in biometric evaluation. Image acquisition and dataset description The proposed framework is evaluated on the CASIA-IrisV3 dataset, utilizing the Interval, Lamp, and Twins subsets, which consist of near-infrared (NIR) iris images captured under controlled and semi-controlled indoor conditions. NIR imaging enhances iris texture visibility by reducing specular reflections and suppressing illumination variability, thereby improving the clarity of discriminative patterns. The dataset comprises three complementary subsets that introduce varying levels of complexity: CASIA-Iris-Interval: 2,639 images from 249 subjects.CASIA-Iris-Lamp: 16,212 images from 411 subjects, incorporating illumination variations.CASIA-Iris-Twins: 3,183 images from 200 subjects, including twin pairs with high inter-subject similarity. CASIA-Iris-Interval: 2,639 images from 249 subjects. CASIA-Iris-Interval: 2,639 images from 249 subjects. CASIA-Iris-Lamp: 16,212 images from 411 subjects, incorporating illumination variations. CASIA-Iris-Lamp: 16,212 images from 411 subjects, incorporating illumination variations. CASIA-Iris-Twins: 3,183 images from 200 subjects, including twin pairs with high inter-subject similarity. CASIA-Iris-Twins: 3,183 images from 200 subjects, including twin pairs with high inter-subject similarity. In total, the evaluation utilizes 22,034 iris images. The inclusion of the Lamp subset introduces illumination variability, while the Twins subset presents a particularly challenging verification scenario due to high inter-subject similarity. This combination increases both the scale and difficulty of the evaluation compared to single-subset protocols. The dataset is formally defined as:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:\mathcal{D}=\left\{\right({I}_{k},{y}_{k}\left)\right\}&\:&\:\end{array}$$\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}$$\:{I}_{k}$$\end{document} denotes the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k$$\end{document}-th iris image and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{k}$$\end{document}represents the corresponding subject identity label. Verification pairs are constructed as:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:\mathcal{P}=\left\{\right({I}_{i},{I}_{j},{y}_{ij}\left)\right\}&\:&\:\end{array}$$\end{document} where 10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{y}_{ij}=\left\{\begin{array}{cc}1,&\:\mathrm{if\:}{y}_{i}={y}_{j}\\\:0,&\:\mathrm{if\:}{y}_{i}\ne\:{y}_{j}\end{array}\right.&\:&\:\end{array}$$\end{document} For each subject with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{s}$$\end{document}samples, the number of genuine comparison pairs is:11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:\frac{{n}_{s}({n}_{s}-1)}{2}&\:&\:\end{array}$$\end{document} All intra-class combinations are used to generate genuine comparisons, resulting in multiple comparison pairs per identity across varying acquisition conditions. Inter-class comparisons are constructed across different subjects to form a large impostor set. This evaluation protocol produces a statistically significant number of genuine and impostor comparison pairs, enabling reliable estimation of False Acceptance Rate (FAR), False Rejection Rate (FRR), and Equal Error Rate (EER), in accordance with standard biometric evaluation practices. In addition to structural metadata, statistical characterization of the dataset is performed to describe image distribution and acquisition properties relevant to biometric analysis, as summarized in Table 1. Table 1Statistical and Acquisition Characteristics of the CASIA-IrisV3 Dataset.ParameterSpecificationDataset NameCASIA-IrisV3Imaging ModalityNear-Infrared (NIR)Subsets UsedInterval, Lamp, TwinsCASIA-Iris-Interval2,639 images/249 subjectsCASIA-Iris-Lamp16,212 images/411 subjectsCASIA-Iris-Twins3,183 images/200 subjectsTotal Images Used22,034Total SubjectsSubject identities may overlap across subsetsImages per SubjectMultiple (subset-dependent)Image Resolution320 × 280 pixelsBit Depth8-bit grayscaleCapture EnvironmentControlled and semi-controlled indoor conditionsDynamic Intensity Range0–255Genuine Comparison Pairs \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sum\:}_{s}\frac{{n}_{s}({n}_{s}-1)}{2}$$\end{document} Impostor Comparison PairsLarge-scale inter-class combinationsEvaluation Protocol1:1 Verification (all intra-class and inter-class comparisons) Table 1Statistical and Acquisition Characteristics of the CASIA-IrisV3 Dataset.ParameterSpecificationDataset NameCASIA-IrisV3Imaging ModalityNear-Infrared (NIR)Subsets UsedInterval, Lamp, TwinsCASIA-Iris-Interval2,639 images/249 subjectsCASIA-Iris-Lamp16,212 images/411 subjectsCASIA-Iris-Twins3,183 images/200 subjectsTotal Images Used22,034Total SubjectsSubject identities may overlap across subsetsImages per SubjectMultiple (subset-dependent)Image Resolution320 × 280 pixelsBit Depth8-bit grayscaleCapture EnvironmentControlled and semi-controlled indoor conditionsDynamic Intensity Range0–255Genuine Comparison Pairs \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sum\:}_{s}\frac{{n}_{s}({n}_{s}-1)}{2}$$\end{document} Impostor Comparison PairsLarge-scale inter-class combinationsEvaluation Protocol1:1 Verification (all intra-class and inter-class comparisons) Statistical and Acquisition Characteristics of the CASIA-IrisV3 Dataset. The evaluation leverages all selected subsets of the CASIA-IrisV3 dataset, incorporating variability in illumination conditions (Lamp) and challenging inter-subject similarity (Twins). This enables a comprehensive assessment of verification performance within near-infrared imaging environments commonly used in biometric systems. The proposed framework is therefore suitable for controlled and semi-controlled deployment scenarios, while extension to cross-dataset evaluation remains a direction for future work. Deep denoising autoencoder-based iris enhancement Although the CASIA-IrisV3 dataset is acquired under near-infrared illumination, residual acquisition-induced perturbations remain present across samples. These include low-amplitude sensor noise, localized intensity fluctuations, and minor structural inconsistencies. While visually subtle, such perturbations can increase intra-class embedding variance and affect verification stability, particularly under challenging inter-class conditions. To improve structural consistency prior to feature extraction, a Deep Denoising Autoencoder (DDAE) is incorporated as an enhancement module. The architecture of the proposed DDAE is illustrated in Fig. 2. Fig. 2Architecture of the Deep Denoising Autoencoder. Architecture of the Deep Denoising Autoencoder. Let the observed iris image be modeled as:12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:I={I}^{*}+\epsilon &\:&\:\end{array}$$\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}$$\:{I}^{*}$$\end{document} denotes the latent clean iris structure and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\epsilon$$\end{document}represents additive acquisition noise. During training, artificial corruption is introduced to improve robustness:13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:\stackrel{\sim}{I}=I+\delta\:&\:&\:\end{array}$$\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}$$\:\delta\:$$\end{document} is stochastic noise sampled from a predefined distribution. The encoder \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{\theta\:}(\cdot\:)$$\end{document}maps the corrupted input into a compact latent representation:14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:\mathbf{h}={f}_{\theta\:}\left(\stackrel{\sim}{I}\right)&\:&\:\end{array}$$\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}$$\:\mathbf{h}\in\:{\mathbb{R}}^{k}$$\end{document} represents the latent embedding of dimensionality \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k$$\end{document}. The decoder\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}_{\phi\:}(\cdot\:)$$\end{document} reconstructs the enhanced image:15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:\widehat{I}={g}_{\phi\:}\left(\mathbf{h}\right)&\:&\:\end{array}$$\end{document} Reconstruction learning is performed by minimizing mean squared error:16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{\mathcal{L}}_{rec}=\frac{1}{HW}\sum\:_{x=1}^{H}\sum\:_{y=1}^{W}{\left(I(x,y)-\widehat{I}(x,y)\right)}^{2}&\:&\:\end{array}$$\end{document} To avoid excessive smoothing of high-frequency iris textures, structural consistency is implicitly preserved by controlling latent compression. Additionally, gradient preservation is monitored using the average gradient magnitude:17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:G=\frac{1}{HW}\sum\:_{x,y}\sqrt{{\left(\frac{\partial\:I}{\partial\:x}\right)}^{2}+{\left(\frac{\partial\:I}{\partial\:y}\right)}^{2}}&\:&\:\end{array}$$\end{document} The enhancement process is designed to suppress stochastic variations while preserving discriminative edge patterns associated with iris microstructures such as crypts and radial striations. From a verification perspective, the effectiveness of the enhancement module is evaluated based on its impact on embedding stability and separability. Specifically, improvements are assessed through intra-class similarity consistency, inter-class separation, and changes in Equal Error Rate (EER) under ablation settings. This ensures that the enhancement stage improves representation quality without distorting identity-specific features. Attention-enhanced U-Net for iris segmentation Following structural enhancement, spatial localization of the iris region is performed to eliminate non-discriminative components such as eyelids, eyelashes, sclera, and periocular background. While enhancement regularizes intensity variations, it does not constrain spatial content. In verification systems, inclusion of background pixels can introduce embedding instability and increase intra-class variance. Therefore, segmentation is incorporated as a spatial regularization step to ensure that downstream representation learning is restricted to anatomically relevant iris regions, as illustrated in Fig. 3. Fig. 3Architecture of the Attention-Enhanced U-Net for Iris Segmentation. Architecture of the Attention-Enhanced U-Net for Iris Segmentation. Let the enhanced image be denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{I}\in\:{\mathbb{R}}^{H\times\:W}$$\end{document}. The segmentation objective is to learn a parametric mapping \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{S}_{{\Psi\:}}(\cdot\:)$$\end{document}, parameterized by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\Psi\:}$$\end{document}, such that:18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:M={S}_{{\Psi\:}}\left(\widehat{I}\right)&\:&\:\end{array}$$\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}$$\:M\in\:\{\mathrm{0,1}{\}}^{H\times\:W}$$\end{document}represents the predicted binary iris mask. The segmented iris region is obtained through element-wise masking:19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{I}_{\mathrm{seg}}=\widehat{I}\odot\:M&\:&\:\end{array}$$\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}$$\:\:\:\odot\:$$\end{document} denotes Hadamard multiplication. Segmentation is implemented using an Attention-Enhanced U-Net architecture consisting of an encoder–decoder structure with skip connections and attention gates. The encoder extracts hierarchical features through convolution and downsampling, while the decoder restores spatial resolution through upsampling. A bottleneck layer captures global contextual information. Standard skip connections directly transfer encoder features to the decoder; however, they may also propagate irrelevant background responses. To address this, attention gates are introduced to selectively filter encoder features before fusion. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{l}$$\end{document} denote the encoder feature map at layer \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:l$$\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}$$\:g$$\end{document} denote the gating signal from the corresponding decoder layer. Before fusion, the gating signal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:g$$\end{document} is upsampled to match the spatial resolution of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{l}$$\end{document}. The attention coefficient at the spatial location \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} is computed as:20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:\alpha\:\left(x\right)=\sigma\:\left({W}^{T}\left(\delta\:({W}_{x}{x}_{l}+{W}_{g}g+b)\right)\right)&\:&\:\end{array}$$\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}$$\:{W}_{x},{W}_{g},W$$\end{document}are learnable parameters, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\delta\:(\cdot\:)$$\end{document}denotes ReLU activation, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:(\cdot\:)$$\end{document} represents the sigmoid function. The refined feature map becomes:21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{\stackrel{\sim}{x}}_{l}=\alpha\:\left(x\right)\cdot\:{x}_{l}&\:&\:\end{array}$$\end{document} The attention coefficients \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:\left(x\right)\in\:\left[\mathrm{0,1}\right]$$\end{document}act as soft spatial masks that emphasize iris-relevant regions while suppressing background structures such as eyelids and sclera. This selective filtering improves boundary localization and reduces the propagation of irrelevant features into the decoding pathway. Since the CASIA-IrisV3 dataset does not provide pixel-level segmentation annotations for all samples, ground-truth masks \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{M}_{g}$$\end{document} are generated using a classical iris localization method based on circular boundary detection. The detected inner (pupil) and outer (limbus) boundaries are used to construct binary iris masks, followed by refinement to remove non-iris regions. These masks serve as pseudo ground truth for supervising the segmentation model. Segmentation performance is optimized using Dice loss:22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{L}_{\mathrm{Dice}}=1-\frac{2\mid\:{M}_{p}\cap\:{M}_{g}\mid\:}{\mid\:{M}_{p}\mid\:+\mid\:{M}_{g}\mid\:}&\:&\:\end{array}$$\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}$$\:{M}_{p}$$\end{document}denotes the predicted segmentation mask and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{M}_{g}$$\end{document}represents the corresponding ground-truth mask. Dice loss is particularly suitable due to the inherent class imbalance between iris and non-iris pixels. From a verification perspective, segmentation is expected to reduce embedding variability caused by background interference. By constraining input regions to the iris, segmentation improves intra-class compactness and stabilizes similarity distributions, thereby supporting more reliable downstream verification. Vision transformer-based global feature embedding In iris biometrics, identity information is distributed across radial and circumferential patterns. The proposed Vision Transformer does not explicitly encode such geometric priors; instead, it learns structural dependencies implicitly through self-attention across spatial patches. The model operates on segmented iris regions using standard patch embeddings and learnable positional encodings, without imposing handcrafted geometric constraints. This design enables flexible and data-driven representation learning while preserving global contextual relationships. The architecture of the Vision Transformer used for feature embedding is illustrated in Fig. 4. Fig. 4Vision Transformer Architecture for Global Iris Embedding. Vision Transformer Architecture for Global Iris Embedding. The segmented iris images are resized to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:224\times\:224$$\end{document} pixels and partitioned into non-overlapping patches 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}, resulting in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:N=196$$\end{document} tokens. Each patch is projected into a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:D=768$$\end{document}-dimensional embedding space. A learnable classification token is appended to the sequence, and positional embeddings are added to retain spatial structure. The transformer encoder consists of 12 layers with 12 attention heads per layer, and the feed-forward network (FFN) has a hidden dimension of 3072. Dropout with a rate of 0.1 is applied within both attention and feed-forward layers to improve generalization. The model is initialized using ImageNet-pretrained weights and fine-tuned end-to-end on the CASIA-IrisV3 dataset (Interval, Lamp, and Twins subsets). Optimization is performed using the Adam optimizer with an initial learning rate of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:1\times\:{10}^{-4}$$\end{document}, weight decay of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:1\times\:{10}^{-5}$$\end{document}, and a batch size of 32. Training is conducted for up to 100 epochs with early stopping based on validation performance, typically converging around 60 epochs. The detailed configuration is summarized in Table 2. Table 2Vision Transformer Configuration and Training Settings.ParameterValueInput resolution224 × 224Patch size16 × 16Number of patches196Embedding dimension768Transformer layers12Attention heads12FFN dimension3072Positional encodingLearnableDropout rate0.1OptimizerAdamLearning rate1 × 10⁻⁴Weight decay1 × 10⁻⁵Batch size32EpochsUp to 100 (early stopping at ~ 60–70 epochs)Training strategyFine-tuning (all layers) Table 2Vision Transformer Configuration and Training Settings.ParameterValueInput resolution224 × 224Patch size16 × 16Number of patches196Embedding dimension768Transformer layers12Attention heads12FFN dimension3072Positional encodingLearnableDropout rate0.1OptimizerAdamLearning rate1 × 10⁻⁴Weight decay1 × 10⁻⁵Batch size32EpochsUp to 100 (early stopping at ~ 60–70 epochs)Training strategyFine-tuning (all layers) Vision Transformer Configuration and Training Settings. Let the segmented iris image be denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{I}_{\mathrm{seg}}\in\:{\mathbb{R}}^{H\times\:W}$$\end{document}. The image is partitioned into non-overlapping patches of size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:P\times\:P$$\end{document}, yielding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:N=\frac{HW}{{P}^{2}}$$\end{document} tokens. Each patch is vectorized and projected into a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:D$$\end{document}-dimensional latent space:23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{x}_{i}={W}_{e}{\hspace{0.17em}}\mathrm{vec}\left({I}_{\mathrm{seg}}^{i}\right)+{b}_{e},i=1,\dots\:,N&\:&\:\end{array}$$\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}$$\:{W}_{e}\in\:{\mathbb{R}}^{D\times\:{P}^{2}}$$\end{document}is a learnable embedding matrix. A learnable classification token \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{\mathrm{cls}}$$\end{document}is appended, and positional embeddings \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{\mathrm{pos}}\in\:{\mathbb{R}}^{(N+1)\times\:D}$$\end{document}are added:24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{Z}_{0}=[{x}_{\mathrm{cls}};{x}_{1};\dots\:{\hspace{0.17em}};{x}_{N}]+{E}_{\mathrm{pos}}&\:&\:\end{array}$$\end{document} The token sequence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Z}_{0}$$\end{document} is processed through \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:L$$\end{document}stacked transformer encoder blocks. Each block consists of a Multi-Head Self-Attention (MHSA) layer followed by a position-wise feed-forward network (FFN), with residual connections and layer normalization applied to ensure stable optimization. Within each attention head, query, key, and value matrices are computed as:25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:Q=Z{W}_{Q},K=Z{W}_{K},V=Z{W}_{V}&\:&\:\end{array}$$\end{document} Self-attention computes contextualized representations as:26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:\mathrm{SA}\left(Z\right)=\mathrm{Softmax}\left(\frac{Q{K}^{{\top\:}}}{\sqrt{{d}_{k}}}\right)V&\:&\:\end{array}$$\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}$$\:{d}_{k}$$\end{document}denotes the key dimensionality. Multi-head attention aggregates complementary representations:27\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:\mathrm{MHSA}\left(Z\right)=\mathrm{Concat}({\mathrm{SA}}_{1},\dots\:,{\mathrm{SA}}_{h}){W}_{O}&\:&\:\end{array}$$\end{document} The feed-forward sublayer applies a nonlinear transformation:28\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:\mathrm{FFN}\left(Z\right)=\sigma\:\left(Z{W}_{1}+{b}_{1}\right){W}_{2}+{b}_{2}&\:&\:\end{array}$$\end{document} After \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:L$$\end{document} encoder layers, the final embedding is extracted from the classification token:29\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:z={Z}_{L}^{\left(\mathrm{cls}\right)}&\:&\:\end{array}$$\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}$$\:z\in\:{\mathbb{R}}^{D}$$\end{document}serves as the global identity representation. From a verification perspective, the self-attention mechanism enables the model to capture global structural relationships across iris patches by modeling interactions among all spatial regions. This reduces reliance on localized features and improves robustness to segmentation imperfections and local texture variations. Formally, attention induces contextualized representation:30\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{z}_{i}^{\mathrm{context}}=\sum\:_{j=1}^{N}{\alpha\:}_{ij}{v}_{j}&\:&\:\end{array}$$\end{document} where attention weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{ij}$$\end{document} represent learned inter-patch dependencies. Such global aggregation improves intra-class compactness and enhances inter-class separability, thereby improving the reliability of cosine similarity in the Siamese verification stage. Siamese similarity learning and verification decision The global embedding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbf{z}\in\:{\mathbb{R}}^{D}$$\end{document}extracted by the Vision Transformer serves as the identity representation for verification. Rather than treating recognition as a closed-set classification task, the framework formulates iris recognition as a metric learning problem within a shared embedding space. A Siamese architecture is employed, where two identical branches with shared parameters process paired iris samples concurrently. The structure is illustrated in Fig. 5. Fig. 5Siamese Network Architecture for Iris Feature Matching and Identity Verification. Siamese Network Architecture for Iris Feature Matching and Identity Verification. For a pair of segmented iris images \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{I}_{i}$$\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}$$\:{I}_{j}$$\end{document}, their embeddings are defined as:31\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{z}_{i}={f}_{\theta\:}\left({I}_{i}\right),{z}_{j}={f}_{\theta\:}\left({I}_{j}\right)&\:&\:\end{array}$$\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}$$\:{f}_{\theta\:}(\cdot\:)$$\end{document} denotes the transformer-based feature extraction function described in the section "Vision transformer-based global feature embedding". Similarity between embeddings is measured using cosine similarity:32\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{S}_{ij}=\frac{{z}_{i}^{{\top\:}}{z}_{j}}{\parallel\:{z}_{i}{\parallel\:}_{2}\parallel\:{z}_{j}{\parallel\:}_{2}}&\:&\:\end{array}$$\end{document} For normalized embeddings (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\parallel\:z{\parallel\:}_{2}=1$$\end{document}), cosine similarity reduces to:33\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{S}_{ij}={z}_{i}^{{\top\:}}{z}_{j}&\:&\:\end{array}$$\end{document} The Euclidean distance between embeddings is defined as:34\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{d}_{ij}=\parallel\:{z}_{i}-{z}_{j}{\parallel\:}_{2}&\:&\:\end{array}$$\end{document} which relates to cosine similarity as:35\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{d}_{ij}^{2}=2(1-{S}_{ij})&\:&\:\end{array}$$\end{document} This equivalence shows that maximizing cosine similarity corresponds to minimizing Euclidean distance for normalized embeddings. To enforce discriminative separation during training, contrastive loss is employed:36\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{L}_{\mathrm{pair}}={y}_{ij}{\hspace{0.17em}}{d}_{ij}^{2}+(1-{y}_{ij}){\hspace{0.17em}}\mathrm{m}\mathrm{a}\mathrm{x}(0,m-{d}_{ij}{)}^{2}&\:&\:\end{array}$$\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}$$\:{y}_{ij}\in\:\left\{\mathrm{0,1}\right\}$$\end{document} denotes the pair label 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} is a predefined margin. This objective encourages compact intra-class clustering while enforcing inter-class separation. From a statistical perspective, verification performance depends on the separation between genuine and impostor similarity distributions. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{g}$$\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}$$\:{\mu\:}_{i}$$\end{document} denote the expected similarity of genuine and impostor pairs, respectively, with corresponding variances \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{g}^{2}$$\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}$$\:{\sigma\:}_{i}^{2}$$\end{document}:37\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu \:_{g} = {\mathbb{E}}\left[ {S_{{ij}} |y_{{ij}} = 1} \right],\mu \:_{i} = {\mathbb{E}}\left[ {S_{{ij}} |y_{{ij}} = 0} \right]$$\end{document}38\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{\sigma\:}_{g}^{2}=\mathrm{V}\mathrm{a}\mathrm{r}({S}_{ij}\mid\:{y}_{ij}=1),{\sigma\:}_{i}^{2}=\mathrm{V}\mathrm{a}\mathrm{r}({S}_{ij}\mid\:{y}_{ij}=0)&\:&\:\end{array}$$\end{document} Effective metric learning increases the separation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{g}-{\mu\:}_{i}$$\end{document} while maintaining low intra-class variance \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{g}^{2}$$\end{document} , thereby reducing distributional overlap and improving verification reliability. During inference, the similarity score is compared against a decision threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\tau\:$$\end{document}:39\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{D}_{ij}=\left\{\begin{array}{cc}1,&\:\mathrm{if\:}{S}_{ij}\ge\:\tau\:\\\:0,&\:\mathrm{otherwise}\end{array}\right.&\:&\:\end{array}$$\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}$$\:{D}_{ij}=1$$\end{document} indicates a genuine match and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{D}_{ij}=0$$\end{document} denotes an impostor pair. The threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\tau\:$$\end{document}is determined using validation-based ROC analysis to balance False Acceptance Rate (FAR) and False Rejection Rate (FRR). By integrating transformer-based global embedding with margin-constrained Siamese optimization, the proposed framework produces discriminative and stable representations suitable for reliable biometric verification. Algorithmic formulation of the proposed transformer–Siamese framework To consolidate the methodological pipeline into a reproducible procedure, the complete training and verification workflow is formalized in Algorithm 1. Algorithm 1End-to-End Iris Verification via Transformer–Siamese Metric Learning. End-to-End Iris Verification via Transformer–Siamese Metric Learning. Algorithm 1 summarizes the end-to-end pipeline, integrating enhancement, segmentation, transformer-based embedding, and metric learning within a unified framework. The training phase optimizes discriminative embedding representations, while the inference phase performs verification through threshold-based similarity comparison. Results and discussions This section presents a comprehensive evaluation of the proposed transformer–Siamese iris verification framework on the CASIA-IrisV3 dataset under a subject-disjoint protocol. The analysis is organized to examine convergence behavior, preprocessing effectiveness, embedding separability, verification performance, and component-wise contributions. The dataset consists of 22,034 iris images acquired under varying conditions. A subject-level partitioning strategy (70% training, 15% validation, and 15% testing) was adopted to eliminate identity overlap across subsets. Verification pairs were generated independently within each partition. For each subject with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{s}$$\end{document} samples, all intra-class combinations were used to construct genuine pairs, while impostor pairs were sampled across different subjects with class balancing applied during training. All images were resized and intensity-normalized prior to processing. The enhancement, segmentation, and embedding modules were trained jointly in an end-to-end manner. The model was optimized using the Adam optimizer with a learning rate of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:1\times\:{10}^{-4}$$\end{document}, weight decay of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:1\times\:{10}^{-5}$$\end{document}, and a batch size of 32. Training was performed for up to 100 epochs with early stopping based on validation loss. To improve generalization, regularization strategies including dropout, weight decay, and data augmentation (intensity variation, rotation, and occlusion simulation) were applied. The decision threshold was determined on the validation set using ROC analysis and fixed prior to evaluation on the test set. All experiments are conducted on a GPU-accelerated platform. The model requires approximately 3 min per epoch during training and achieves an average inference latency of approximately 12.6 ms per verification pair. All experiments are conducted using fixed random seeds to ensure reproducibility. Convergence behavior and optimization stability The optimization dynamics of the proposed framework were evaluated at 10-epoch intervals to assess loss convergence and generalization behavior. Table 3Training and Validation Convergence Statistics.EpochTrain LossVal LossTrain Acc (%)Val Acc (%)10.8650.88260.458.9100.6120.63873.872.1200.4380.46283.581.7300.3190.34189.487.6400.2450.26293.291.5500.1980.21495.193.6600.1760.19196.394.8700.1710.18896.694.7 Table 3Training and Validation Convergence Statistics.EpochTrain LossVal LossTrain Acc (%)Val Acc (%)10.8650.88260.458.9100.6120.63873.872.1200.4380.46283.581.7300.3190.34189.487.6400.2450.26293.291.5500.1980.21495.193.6600.1760.19196.394.8700.1710.18896.694.7 Training and Validation Convergence Statistics. Table 3 shows that both training and validation losses decrease consistently during early training, accompanied by a steady increase in accuracy. Convergence is reached at approximately epoch 60, where validation loss stabilizes (≈ 0.191) and accuracy approaches saturation. The gap between training and validation performance remains small throughout training, indicating stable generalization without evident overfitting. Beyond epoch 60, improvements in validation accuracy are marginal, and early stopping is therefore applied at this stage. Fig. 6Training and Validation Accuracy Across Epochs. Training and Validation Accuracy Across Epochs. Fig. 7Training and Validation Loss Across Epochs. Training and Validation Loss Across Epochs. Figure 6 illustrates the progression of training and validation accuracy, while Fig. 7 shows the corresponding loss curves. Both figures exhibit smooth and consistent trends, supporting stable optimization behavior. Enhancement and segmentation validation This section evaluates the effectiveness of the enhancement and segmentation modules in improving input quality and isolating the iris region prior to feature extraction. The analysis is conducted on the validation subset of the CASIA-IrisV3 Interval, Lamp, and Twins subsets using quantitative image quality metrics and segmentation accuracy measures. Enhancement performance is assessed using entropy (H), gradient energy density (GED), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM). These metrics capture complementary aspects of image quality, including textural richness, edge strength, noise reduction, and structural preservation. Table 4Enhancement Quality Metrics.MetricRaw ImageEnhanced ImageEntropy (H)6.08 ± 0.286.54 ± 0.24Gradient Energy Density0.179 ± 0.0230.236 ± 0.019PSNR (dB)27.9 ± 2.431.6 ± 1.8SSIM0.882 ± 0.0310.934 ± 0.022 Table 4Enhancement Quality Metrics.MetricRaw ImageEnhanced ImageEntropy (H)6.08 ± 0.286.54 ± 0.24Gradient Energy Density0.179 ± 0.0230.236 ± 0.019PSNR (dB)27.9 ± 2.431.6 ± 1.8SSIM0.882 ± 0.0310.934 ± 0.022 Enhancement Quality Metrics. As shown in Table 4, enhancement increases entropy from 6.08 to 6.54, indicating improved textural richness. Gradient energy density increases substantially, reflecting stronger edge representation. Improvements in PSNR and SSIM indicate that noise is reduced while structural fidelity is preserved. Segmentation performance is evaluated using overlap-based and boundary-aware metrics. Table 5Segmentation Boundary Localization Performance.MetricMean ± StdDice Similarity Coefficient (DSC)0.957 ± 0.016Intersection over Union (IoU)0.918 ± 0.019Precision0.949 ± 0.017Recall0.964 ± 0.014Boundary F1-score0.939 ± 0.015 Table 5Segmentation Boundary Localization Performance.MetricMean ± StdDice Similarity Coefficient (DSC)0.957 ± 0.016Intersection over Union (IoU)0.918 ± 0.019Precision0.949 ± 0.017Recall0.964 ± 0.014Boundary F1-score0.939 ± 0.015 Segmentation Boundary Localization Performance. Table 5 shows that the segmentation module achieves high overlap accuracy, with a Dice coefficient of 0.957 and an IoU of 0.918. The balance between precision and recall indicates accurate foreground localization with minimal background inclusion. The boundary F1-score further reflects the precise delineation of iris contours. Fig. 8Stage-wise Iris Processing Results: a Preprocessed Image, b Ground Truth mask, c Predicted Mask, d Attention map, e Overlay. Stage-wise Iris Processing Results: a Preprocessed Image, b Ground Truth mask, c Predicted Mask, d Attention map, e Overlay. Figure 8 presents qualitative results of the preprocessing pipeline, including enhanced images, predicted masks, attention maps, and overlay. The results show that the model reduces background interference while preserving discriminative iris structures, supporting reliable input preparation for subsequent feature extraction. Embedding space separability and discriminative analysis This section analyzes the discriminative quality of the learned embedding space by examining the separation between genuine and impostor similarity distributions. The evaluation is conducted on the validation subset of the CASIA-IrisV3 Interval, Lamp, and Twins subsets after convergence. Table 6Statistical Characterization of Cosine Similarity Distributions.StatisticGenuine PairsImpostor PairsMean Similarity (µ)0.894 ± 0.0260.349 ± 0.052Median0.8970.341Variance (σ²)0.000680.002705th Percentile0.8320.23895th Percentile0.9380.489Mean Separation (µg − µi)0.545–Fisher Discriminant Ratio (FDR)121.3– Table 6Statistical Characterization of Cosine Similarity Distributions.StatisticGenuine PairsImpostor PairsMean Similarity (µ)0.894 ± 0.0260.349 ± 0.052Median0.8970.341Variance (σ²)0.000680.002705th Percentile0.8320.23895th Percentile0.9380.489Mean Separation (µg − µi)0.545–Fisher Discriminant Ratio (FDR)121.3– Statistical Characterization of Cosine Similarity Distributions. Table 6 shows a mean separation of 0.545 between genuine and impostor similarity distributions, indicating strong separation in the embedding space. Compared to subset-based evaluation, the slightly reduced separation and increased variance reflect the inclusion of more challenging samples, such as illumination variations and highly similar iris patterns. The Fisher Discriminant Ratio (121.3) further indicates a large inter-class separation relative to intra-class dispersion. The relatively compact genuine distribution, combined with a broader impostor distribution, suggests stable identity clustering despite increased dataset complexity. Fig. 9Embedding Space Separability Analysis: a Cosine similarity distribution; b t-SNE embeddings visualization. Embedding Space Separability Analysis: a Cosine similarity distribution; b t-SNE embeddings visualization. Figure 9(a) presents the distribution of cosine similarity scores for genuine and impostor pairs, showing limited overlap between the two distributions. The decision threshold (τ = 0.69) lies within the separation region, supporting stable verification performance. Figure 9(b) illustrates a t-SNE visualization of the embedding space across multiple identity classes. Distinct clusters are observed with limited inter-class overlap, indicating that the learned representations preserve identity-specific structure. Fig. 10Visualization of Vision Transformer Attention a Input Iris Image, b Attention map, c Overlay. Visualization of Vision Transformer Attention a Input Iris Image, b Attention map, c Overlay. To further examine the feature extraction process, Fig. 10 presents attention visualizations from the Vision Transformer. The attention maps highlight discriminative iris texture regions while suppressing responses to non-iris areas. This suggests that the model captures spatially distributed structural patterns rather than relying solely on localized features. Pattern matching and verification performance Verification performance is evaluated on the independent test subset of the CASIA-IrisV3 Interval, Lamp, and Twins subsets using the cosine similarity decision rule defined in the section "Siamese similarity learning and verification decision". The operating threshold (τ) is determined on the validation set via ROC analysis and fixed prior to testing. Fig. 11Visual comparison of genuine and impostor embedding alignment in the pattern matching stage. Visual comparison of genuine and impostor embedding alignment in the pattern matching stage. Figure 11 illustrates qualitative probe–gallery comparisons at the pattern matching stage. The upper row represents a genuine pair exhibiting strong structural correspondence, while the lower row shows an impostor pair with weaker alignment. Highlighted regions indicate discriminative texture areas contributing to similarity computation in the embedding space. Table 7Test-Set Verification Performance.MetricValueEqual Error Rate (EER)2.34%Area Under ROC Curve (AUC)0.987True Acceptance Rate (TAR) @ FAR = 10⁻³≈ 95%True Acceptance Rate (TAR) @ FAR = 10⁻²98.1%False Non-Match Rate (FNMR) @ FAR = 10⁻³≈ 5.0%Verification Accuracy97.82%False Acceptance Rate (FAR) @ τ = 0.692.41%False Rejection Rate (FRR) @ τ = 0.692.28% Table 7Test-Set Verification Performance.MetricValueEqual Error Rate (EER)2.34%Area Under ROC Curve (AUC)0.987True Acceptance Rate (TAR) @ FAR = 10⁻³≈ 95%True Acceptance Rate (TAR) @ FAR = 10⁻²98.1%False Non-Match Rate (FNMR) @ FAR = 10⁻³≈ 5.0%Verification Accuracy97.82%False Acceptance Rate (FAR) @ τ = 0.692.41%False Rejection Rate (FRR) @ τ = 0.692.28% Test-Set Verification Performance. As summarized in Table 7, the proposed framework achieves an Equal Error Rate (EER) of 2.34% and an Area Under the ROC Curve (AUC) of 0.987, indicating strong separation between genuine and impostor distributions. The close alignment between FAR (2.41%) and FRR (2.28%) reflects consistent threshold calibration. Verification accuracy (97.82%) is reported as a secondary metric, while EER and ROC-based measures serve as the primary evaluation criteria. In addition to threshold-independent metrics, deployment-oriented performance is evaluated using TAR at fixed FAR levels. The model achieves approximately 95% TAR at FAR = 10⁻³ and 98.1% at FAR = 10⁻², indicating stable performance under stricter operating conditions. Fig. 12Receiver Operating Characteristic (ROC) Curve on the Test set. Receiver Operating Characteristic (ROC) Curve on the Test set. Figure 12 illustrates the trade-off between False Positive Rate (FPR) and True Positive Rate (TPR) across varying decision thresholds. The operating point corresponding to the Equal Error Rate (EER = 2.34%) is indicated, where FPR ≈ 2.34% and TPR ≈ 97.66%. The Area Under the Curve (AUC = 0.987) reflects strong discriminative capability. Fig. 13Detection Error Trade-off (DET) Curve of the Proposed Verification Framework. Detection Error Trade-off (DET) Curve of the Proposed Verification Framework. Figure 13 shows the relationship between False Acceptance Rate (FAR) and False Rejection Rate (FRR) across varying decision thresholds on a logarithmic scale. The operating point corresponding to the Equal Error Rate (EER = 2.34%) is highlighted, with FAR ≈ 2.41% and FRR ≈ 2.28%, consistent with Table 7. Component-wise ablation analysis To quantify the contribution of individual components, ablation experiments are conducted under identical training, validation, and threshold calibration protocols using the CASIA-IrisV3 Interval, Lamp, and Twins subsets. Each variant preserves subject-disjoint partitioning to ensure fair comparison. Performance is evaluated in terms of Equal Error Rate (EER), verification accuracy, Area Under the ROC Curve (AUC), and embedding separation. Table 8Component-wise Performance and Embedding Separation Analysis.Model VariantEER (%)ΔEER (%)Accuracy (%)AUCMean Separation (µg − µi)Full Proposed Model2.34–97.820.9870.545Without DDAE3.41+ 45.7%95.380.9790.472Without Segmentation4.02+ 71.8%94.610.9710.438CNN Backbone2.98+ 27.4%95.940.9820.501 Table 8Component-wise Performance and Embedding Separation Analysis.Model VariantEER (%)ΔEER (%)Accuracy (%)AUCMean Separation (µg − µi)Full Proposed Model2.34–97.820.9870.545Without DDAE3.41+ 45.7%95.380.9790.472Without Segmentation4.02+ 71.8%94.610.9710.438CNN Backbone2.98+ 27.4%95.940.9820.501 Component-wise Performance and Embedding Separation Analysis. Table 8 shows that removing the enhancement module increases EER to 3.41%, indicating that residual noise affects embedding consistency. Eliminating segmentation results in the largest degradation (EER = 4.02%), highlighting the importance of spatial localization in reducing periocular interference. Replacing the transformer backbone with a CNN leads to reduced embedding separation and increased EER, suggesting that global contextual modeling contributes to improved representation quality. Table 9Threshold Sensitivity Analysis (EER %).Model VariantEER @ τ − 0.02EER @ τEER @ τ + 0.02Full Proposed Model2.582.342.49Without DDAE3.923.413.76Without Segmentation4.614.024.38CNN Backbone3.212.983.15 Table 9Threshold Sensitivity Analysis (EER %).Model VariantEER @ τ − 0.02EER @ τEER @ τ + 0.02Full Proposed Model2.582.342.49Without DDAE3.923.413.76Without Segmentation4.614.024.38CNN Backbone3.212.983.15 Threshold Sensitivity Analysis (EER %). Table 9 shows that the proposed model exhibits low sensitivity to threshold perturbations (approximately ± 0.15% variation in EER), indicating stable similarity distributions and consistent decision boundaries. In contrast, ablated variants demonstrate higher sensitivity, reflecting increased overlap between genuine and impostor distributions. Table 10Comprehensive Ablation Analysis of Design Components and Interaction Effects.ConfigurationLossAttentionViTEER (%)ΔEER (%)AUCSeparationFull ModelContrastive✓Base2.34–0.9870.545No Contrastive LossBCE✓Base3.12+ 33.3%0.9810.498No Attention GateContrastive✗Base3.08+ 31.6%0.9800.486No Contrastive + No AttentionBCE✗Base3.89+ 66.2%0.9730.442ViT-SmallContrastive✓Small2.91+ 24.4%0.9810.507ViT-LargeContrastive✓Large2.29−2.1%0.9880.552 Table 10Comprehensive Ablation Analysis of Design Components and Interaction Effects.ConfigurationLossAttentionViTEER (%)ΔEER (%)AUCSeparationFull ModelContrastive✓Base2.34–0.9870.545No Contrastive LossBCE✓Base3.12+ 33.3%0.9810.498No Attention GateContrastive✗Base3.08+ 31.6%0.9800.486No Contrastive + No AttentionBCE✗Base3.89+ 66.2%0.9730.442ViT-SmallContrastive✓Small2.91+ 24.4%0.9810.507ViT-LargeContrastive✓Large2.29−2.1%0.9880.552 Comprehensive Ablation Analysis of Design Components and Interaction Effects. Table 10 shows that removing contrastive learning or attention gating individually leads to performance degradation, while their joint removal results in a larger increase in EER, indicating complementary contributions. Increasing transformer capacity yields a marginal improvement, whereas reducing capacity leads to a noticeable performance drop. These results suggest that the selected configuration provides a balanced trade-off between verification performance and computational efficiency. Robustness and threshold sensitivity analysis To further evaluate the stability of the proposed verification framework, robustness is analyzed with respect to decision threshold variations and similarity distribution characteristics. The evaluation is conducted on the CASIA-IrisV3 Interval, Lamp, and Twins subsets under the same subject-disjoint protocol. The threshold sensitivity results (Table 9) show that the proposed model exhibits low variation in performance under small perturbations of the decision threshold (± 0.02). The resulting change in EER remains within approximately ± 0.15%, indicating stable separation between genuine and impostor similarity distributions. In contrast, ablated configurations demonstrate higher sensitivity to threshold variations, reflecting increased overlap between similarity distributions. This observation is consistent with the reduced embedding separation reported in the section "Component-wise ablation analysis" and indicates that the integrated design contributes to more stable decision boundaries. Additionally, robustness is supported by the consistent alignment between validation and test performance, as well as the small train–validation gap observed during convergence (Section "Convergence behavior and optimization stability"). These observations indicate stable model behavior across operating conditions without requiring frequent adjustment of the decision threshold. Comparative analysis with state-of-the-art methods To contextualize the performance of the proposed transformer–Siamese framework, comparisons are made with representative iris recognition approaches reported on the CASIA-IrisV3 dataset, as summarized in Table 11. Table 11Comparison with Representative Iris Recognition Methods.ApproachDataset UsedPerformanceEnhanced UNet + + segmentation architecture12 CASIA-IrisV3 96.21% Complex-Valued CNN11 CASIA-IrisV3 95.00% Hybrid biometric authentication31 CASIA-IrisV3 95.07% CapsuleNet with enhanced routing28 CASIA-V4 Lamp 96.30%, EER: 1.241%Proposed Framework CASIA-IrisV3 EER: 2.34%, AUC: 0.987, Acc†: 97.82%Accuracy is computed at the operating threshold corresponding to the Equal Error Rate (EER). The proposed framework is primarily evaluated using verification metrics (EER and AUC), whereas prior works often report identification-based accuracy under different evaluation protocols. Therefore, direct numerical comparison should be interpreted with caution. Table 11Comparison with Representative Iris Recognition Methods.ApproachDataset UsedPerformanceEnhanced UNet + + segmentation architecture12 CASIA-IrisV3 96.21% Complex-Valued CNN11 CASIA-IrisV3 95.00% Hybrid biometric authentication31 CASIA-IrisV3 95.07% CapsuleNet with enhanced routing28 CASIA-V4 Lamp 96.30%, EER: 1.241%Proposed Framework CASIA-IrisV3 EER: 2.34%, AUC: 0.987, Acc†: 97.82%Accuracy is computed at the operating threshold corresponding to the Equal Error Rate (EER). The proposed framework is primarily evaluated using verification metrics (EER and AUC), whereas prior works often report identification-based accuracy under different evaluation protocols. Therefore, direct numerical comparison should be interpreted with caution. Comparison with Representative Iris Recognition Methods. Accuracy is computed at the operating threshold corresponding to the Equal Error Rate (EER). The proposed framework is primarily evaluated using verification metrics (EER and AUC), whereas prior works often report identification-based accuracy under different evaluation protocols. Therefore, direct numerical comparison should be interpreted with caution. Most prior works report identification-based performance metrics, primarily classification accuracy, while the proposed framework is formulated as a verification problem and evaluated using threshold-independent and deployment-relevant metrics such as Equal Error Rate (EER) and Area Under the ROC Curve (AUC). This difference in problem formulation limits direct numerical comparability. To maintain a consistent evaluation setting, the proposed framework is assessed on the complete CASIA-IrisV3 dataset, incorporating the Interval, Lamp, and Twins subsets under a subject-disjoint protocol. This setup introduces variability in illumination conditions and inter-subject similarity, resulting in a more challenging evaluation compared to subset-based or constrained protocols used in some prior studies. Recent transformer-based approaches, such as IrisFormer29, demonstrate the effectiveness of global self-attention for iris representation learning. However, these methods are typically evaluated on different datasets (e.g., CASIA-Thousand, ND-IRIS) or under alternative protocols and metrics. Therefore, they are discussed qualitatively rather than included in the tabulated comparison to avoid misleading conclusions. In contrast, the proposed framework integrates enhancement, attention-guided segmentation, transformer-based embedding, and Siamese metric learning within a unified verification pipeline. This design supports improved structural consistency and enhanced embedding separability under varying acquisition conditions. Overall, the proposed method demonstrates strong verification performance within the evaluated setting while maintaining a balance between accuracy, robustness, and computational efficiency. Computational efficiency and experimental discussion Computational efficiency The computational characteristics of the proposed framework were evaluated under the same hardware configuration used for model training and testing. Profiling was performed for both training and inference stages to assess practical deployment feasibility. Table 12Computational Profiling of the Proposed Framework.CategoryMetricMeasured ValueSettingModel ComplexityTrainable Parameters32.4 MFinal ModelModel Size (Disk)124 MBFP32 WeightsFLOPs (per inference)~ 6.8 GFLOPsSingle Forward PassTrainingConvergence Epochs60Early StoppingTime per Epoch~ 3 minBatch size 32Total Training Time~ 3.6 hFull TrainingOptimizerAdamLR = 1 × 10⁻⁴InferenceLatency per Pair~ 12.6 msBatch size 1Throughput~ 79 pairs/secSingle GPUBatch Throughput~ 420 pairs/secBatch size 32Resource UsagePeak GPU Memory (Training)~ 4.8 GBMixed PrecisionPeak GPU Memory (Inference)~ 2.3 GBBatch size 1CPU Utilization~ 35%Data LoadingDeploymentFrameworkPyTorchGPU-acceleratedHardwareNVIDIA RTX (12 GB)Single GPU Table 12Computational Profiling of the Proposed Framework.CategoryMetricMeasured ValueSettingModel ComplexityTrainable Parameters32.4 MFinal ModelModel Size (Disk)124 MBFP32 WeightsFLOPs (per inference)~ 6.8 GFLOPsSingle Forward PassTrainingConvergence Epochs60Early StoppingTime per Epoch~ 3 minBatch size 32Total Training Time~ 3.6 hFull TrainingOptimizerAdamLR = 1 × 10⁻⁴InferenceLatency per Pair~ 12.6 msBatch size 1Throughput~ 79 pairs/secSingle GPUBatch Throughput~ 420 pairs/secBatch size 32Resource UsagePeak GPU Memory (Training)~ 4.8 GBMixed PrecisionPeak GPU Memory (Inference)~ 2.3 GBBatch size 1CPU Utilization~ 35%Data LoadingDeploymentFrameworkPyTorchGPU-acceleratedHardwareNVIDIA RTX (12 GB)Single GPU Computational Profiling of the Proposed Framework. Table 12 shows that the model converges within 60 epochs under early stopping and maintains moderate computational complexity relative to its architecture. During inference, the framework achieves low latency (approximately 12.6 ms per pair) and high throughput, supporting real-time verification scenarios. Memory usage remains within practical limits for single-GPU deployment. Experimental discussion The experimental results indicate that the proposed framework achieves consistent verification performance under a multi-subset evaluation setting using the CASIA-IrisV3 Interval, Lamp, and Twins subsets. The enhancement stage improves input quality by increasing textural representation, while the segmentation module enables accurate spatial isolation of the iris region prior to feature extraction. The learned embedding space exhibits clear separation between genuine and impostor similarity distributions, supporting reliable verification performance. On the independent test set, the framework achieves an Equal Error Rate (EER) of 2.34% and an Area Under the ROC Curve (AUC) of 0.987 at a fixed decision threshold. The low sensitivity to threshold variations further indicates stable decision boundaries under the evaluated conditions. Ablation analysis shows that segmentation contributes the most significant performance improvement, followed by enhancement and transformer-based feature extraction. Replacing the transformer backbone with a convolutional network reduces embedding separability, highlighting the role of global contextual modeling in representation learning. From a computational perspective, the framework maintains moderate complexity while achieving practical inference performance (approximately 12.6 ms per pair), indicating suitability for near-infrared biometric verification systems under controlled and semi-controlled conditions. The consistency between validation and test performance, along with stable embedding separability, indicates reliable model behavior within the evaluated setting. However, the current evaluation is limited to near-infrared imaging using the CASIA-IrisV3 Interval, Lamp, and Twins subsets. While these subsets introduce variability in illumination and inter-subject similarity, cross-dataset evaluation has not been performed. Therefore, further validation on additional datasets and imaging conditions is required to assess generalization beyond the current experimental scope. Conclusion and future work This work presents a unified transformer–Siamese framework for iris verification that integrates image enhancement, attention-guided segmentation, and global feature extraction within a metric-learning formulation. The proposed approach is evaluated under a subject-disjoint protocol on the CASIA-IrisV3 Interval, Lamp, and Twins subsets. The framework achieves an Equal Error Rate (EER) of 2.34% and an Area Under the ROC Curve (AUC) of 0.987, supported by clear separation between genuine and impostor similarity distributions. Ablation analysis demonstrates that spatial localization, enhancement, and transformer-based representation learning contribute to improved verification performance. The framework maintains moderate computational complexity while achieving efficient inference, providing a balance between verification performance and computational cost. These findings are established within the scope of near-infrared iris recognition and may not directly generalize to visible-spectrum or cross-sensor scenarios. Future work will focus on extending the framework to cross-dataset and cross-sensor evaluation to assess robustness under diverse acquisition conditions. Exploration of lightweight transformer architectures and model compression techniques may further improve scalability. Additionally, evaluating performance under challenging conditions such as occlusion, pupil dilation variation, and motion blur will help assess applicability in less constrained biometric settings.