Spatially context aware multilevel color image segmentation using a Hybrid Artificial Hummingbird and Great Deluge mechanism Authors: Tirumalasetti Supraja, Kankanala Srinivas Published: 2026 ABSTRACT Multilevel image thresholding is an important segmentation technique that partitions an image into meaningful regions in applications such as object recognition, medical imaging, and satellite image analysis. However, conventional techniques are limited by their poor sensitivity to initial conditions, lack of sufficient spatial contextual information, and slow convergence. The proposed method uses a novel hybrid model based on the Artificial Hummingbird Algorithm (AHA) to address the limitations of existing approaches. Here, Latin Hypercube, Sobol, Halton, and Sierpinski strategies are used during the population initialization phase to improve population diversity and search space coverage. This improves the exploration capability of the algorithm and supports better search space. The proposed methodology also uses spatial contextual information to improve the quality of segmentation. It also incorporates a relationship between neighboring pixels in order to retain more structure and enhance the visual performance of the output. For the exploitation phase, the Great Deluge Algorithm (GDA) is utilized as the optimization algorithm. Furthermore, the use of GDA serves as an adaptive acceptance function which reduces the chance of getting stuck during the search process. Minimum Cross Entropy Measure (MCEM) is used as the objective function to obtain optimal threshold values. Different evaluation metrics have been used to compare the results of the proposed method with other existing metaheuristic algorithms. The code is available at https://github.com/suprajatirumalasetti/AHA_GDA_Image_Segmentation_Code. Introduction The process of image segmentation has a major role to play in the field of computer vision and image processing, since it helps in partitioning the image into several meaningful components for further processing. The process of image segmentation holds significance for the analysis and interpretation of various kinds of images such as medical images, satellite images, natural images, and so forth. This method has been used in medical imaging as well as industrial inspection1. The optimization of threshold values is an NP-hard problem; therefore, exhaustive search becomes computationally impractical. In this context, bio inspired metaheuristic algorithms are useful for handling nonlinear and multimodal search spaces2. The foraging behavior of hummingbirds has recently been modeled in Artificial Hummingbird Algorithm (AHA) proposed by Zhao et al.3. AHA balances local exploitation and global exploration through guided foraging, territorial foraging, migration foraging, and a memory mechanism that records visited food sources. Despite its effectiveness, AHA could be subject to some weaknesses common to population based approaches, such as early convergence, dependency on the initial conditions, and local optimum. Two critical techniques that can enhance AHA include improvement in the criterion of acceptance for the candidate solution and better population coverage through initialization. The AHA has been combined with Great Deluge Algorithm (GDA)4 that utilizes the water level concept for accepting candidate solution. By using this, the system can accept non improving solutions with control due to change in water level, thus making the system efficient in overcoming local minima. Like Simulated Annealing (SA), GDA has been proved to work well in finding near optimal solution in large scale optimization problem. Also, like randomization, random initialization can cause candidates to be initialized at limited regions of the search space, resulting in less diversity in population as well as search space. Therefore, four different types of initialization scheme have been employed in this proposed method: Latin Hypercube Sampling (LHS)5, Sobol, Halton, and Sierpinski initialization6. The performance of multilevel thresholding largely depends upon the fitness function chosen. Entropy criteria is popularly used in segmentation and Minimum Cross Entropy Measure (MCEM) works well since it minimizes the distributional distance between the two distributions7. Unlike variance based OTSU8 criterion, MCEM relies on distributional criteria which make it more suitable for multilevel thresholding. In contrast to conventional histograms that are mainly focused on pixel intensities, this paper presents an approach using energy curves to assist in finding the best possible thresholds9,10. An energy curve can help to smooth the histogram and account for spatial context by offering a continuous version of the histogram, leading to better results concerning the detection of the threshold region and therefore avoiding structure loss, maintaining segmentations borders, and providing a better balance of exploration and exploitation. We offer a mixed AHA approach with MCEM used as a fitness function, GDA as the accepting criteria and four initialization approaches. Our framework is tested using two benchmark datasets: BSDS500 for natural images11 and NASA Earth Observatory dataset (NASA) for satellite images12. The key contributions of our approach are:Energy curve is obtained by means of smoothing of histograms and taking into account spatial context; the continuous landscape obtained defines the threshold locations, retains structural information, and stabilizes the segmentation processThe proposed method combines AHA and GDA methods for segmentation using thresholding at multiple levels; this combination increases exploration capability and avoids early convergence.There are four strategies for initialization of algorithm parameters, namely Latin Hypercube sampling, Sobol, Halton, and Sierpinski used to improve solution diversity and optimization performance.The proposed framework is validated on the BSDS500 dataset for natural images and the NASA Earth Observatory dataset for satellite images, with MCEM used as the objective function.Extensive experiments using Peak Signal to Noise Ratio (PSNR)13, Structural Similarity Index Measure (SSIM)14, Feature Similarity Index Measure (FSIM)15, Quality Index based on Local Variance (QILV)16, Correlation Coefficient (CORR)17, Edge Preservation Index (EPI)18, and Mutual Information Factor (MIF)19, along with statistical significance tests, show that the proposed method provides reliable and competitive performance compared with recent metaheuristic algorithms.The rest of this paper is organized as follows. Section “Literature review” reviews the relevant literature on multilevel thresholding techniques, optimization based segmentation and metaheuristic algorithms. Section “Proposed work” describes the proposed algorithm for multilevel thresholding, including a pixel wise energy curve to catch local intensity variations. This gives more details to the integration of AHA with MCEM, population initialization strategies for enhanced diversity and local refinement via GDA. Section “Results and discussion” presents the experimental results and corresponding discussion, while section “Conclusion” concludes the paper and highlights directions for future research. Energy curve is obtained by means of smoothing of histograms and taking into account spatial context; the continuous landscape obtained defines the threshold locations, retains structural information, and stabilizes the segmentation process Energy curve is obtained by means of smoothing of histograms and taking into account spatial context; the continuous landscape obtained defines the threshold locations, retains structural information, and stabilizes the segmentation process The proposed method combines AHA and GDA methods for segmentation using thresholding at multiple levels; this combination increases exploration capability and avoids early convergence. The proposed method combines AHA and GDA methods for segmentation using thresholding at multiple levels; this combination increases exploration capability and avoids early convergence. There are four strategies for initialization of algorithm parameters, namely Latin Hypercube sampling, Sobol, Halton, and Sierpinski used to improve solution diversity and optimization performance. There are four strategies for initialization of algorithm parameters, namely Latin Hypercube sampling, Sobol, Halton, and Sierpinski used to improve solution diversity and optimization performance. The proposed framework is validated on the BSDS500 dataset for natural images and the NASA Earth Observatory dataset for satellite images, with MCEM used as the objective function. The proposed framework is validated on the BSDS500 dataset for natural images and the NASA Earth Observatory dataset for satellite images, with MCEM used as the objective function. Extensive experiments using Peak Signal to Noise Ratio (PSNR)13, Structural Similarity Index Measure (SSIM)14, Feature Similarity Index Measure (FSIM)15, Quality Index based on Local Variance (QILV)16, Correlation Coefficient (CORR)17, Edge Preservation Index (EPI)18, and Mutual Information Factor (MIF)19, along with statistical significance tests, show that the proposed method provides reliable and competitive performance compared with recent metaheuristic algorithms. Extensive experiments using Peak Signal to Noise Ratio (PSNR)13, Structural Similarity Index Measure (SSIM)14, Feature Similarity Index Measure (FSIM)15, Quality Index based on Local Variance (QILV)16, Correlation Coefficient (CORR)17, Edge Preservation Index (EPI)18, and Mutual Information Factor (MIF)19, along with statistical significance tests, show that the proposed method provides reliable and competitive performance compared with recent metaheuristic algorithms. Literature review Image segmentation is the most vital preprocessing stage in image processing where an image is divided into significant parts also known as image thresholding. Over the past few decades, numerous thresholding techniques have been introduced in the literature1. Threshold segmentation is the most commonly used method for image segmentation20. Accurate and robust segmentation requires consistent pixel attributes throughout each region21. Threshold based segmentation approaches in image segmentation are often classified into two types22. Bilevel thresholding is a fundamental technique that segments an image into two separate classes by determining a single optimal threshold value, it is insufficient for satellite image segmentation23. On the other hand, multilevel thresholding partitions the image into multiple segments by locating several threshold values. Multilevel thresholding is a straightforward and efficient technique in image classification nevertheless, the cost of computation is expensive as the levels of threshold rise. Satellite images have a variety of distinct and complicated objects: the territory, plants, buildings, mountain, urban regions, etc. These areas are rather difficult to differentiate in the poor quality of the image and illumination. Bilevel thresholding is thus inadequate where a satellite image segmentation23. Over the past few decades, literature has reported over hundreds of ways in which the global thresholding can be performed as a part of image segmentation processes24. The techniques of thresholding are classified as parametric and nonparametric models. Nonparametric models seek the thresholds, which maximize an objective function, introduced by Otsu8 and Kapur25. In 1988, Sahoo concluded that the best thresholding image segmentation method was as illustrated by Otsu and gained popularity in literature24,26,27. Nonetheless, the intensive search conditions of the Otsu algorithm are relatively time consuming, and thus have been replaced by various metaheuristic (optimization) algorithms to overcome this problem. Inspired by the natural life systems, the predominant metaheuristic algorithms include the Genetic Algorithm (GA)28, Differential Evolution (DE)29, the Artificial Bee Colony (ABC)30, Particle Swarm Optimization (PSO)31, and the Moth Flame Optimization (MFO)32, Aquila Optimization (AO)33, Equilibrium Optimization (EO)34, and Whale Optimization Algorithm (WOA)35. The no free lunch theorem (NFLT) states that in every problem no single optimization algorithm can be superior to any other algorithm36. From the above survey, classical histogram based methods like the Otsu8 and Kapur often fail to capture spatial contextual information, leading to inconsistent segmentation performance in multilevel thresholding. Recent studies have proposed energy curve model that averages the histogram and reflects the information of the neighboring areas of the image and thus provides a more accurate representation of the image structures37,38. Even if the thresholding algorithm using the metaheuristic is useful in simplifying images, retaining spatial context information becomes difficult. The theoretical basis for doing so is possible through spatial processing methods, which include the mathematics of image processing based on diffusions39, as well as edge preservation of structures used in focusing in multiple images40. Taking cues from such spatial preservation concepts, the energy curve model suggested in our paper takes into consideration the interaction among neighboring pixels. Energy curve driven segmentation approaches were explored in37,38, and employing Electro Magnetism Optimization (EMO) and the Smell Agent Optimizer to enhance multilevel threshold selection and segmentation performance41. Although there are significant improvements in the multilevel thresholding and metaheuristic based image segmentation, none of the studies has been done to synthesize structured initialization strategies and energy curve modeling particularly when it involves RGB image segmentation. Furthermore, the two factors have not been combined into a hybrid architecture where the AHA and the GDA acceptance mechanism. Although Hybrid framework based on energy curves from AHA42 and Hybrid CPA43 have significantly enhanced the capability of multilevel color image segmentation; however they still do not include the integration of GDA, and the use of different heterogeneity initialization strategies. To address this gap, this paper proposes a new standard of segmentation that (i) employs four structured patterns of initializations (Latin Hypercube, Sobol, Halton and Sierpinski) to enhance the diversity and stability of the population (ii) uses energy curve modelling to increase the contextual fidelity of threshold values and (iii) extends the process to RGB images, thereby guaranteeing a strong and stable performance. The behavior of the suggested approach was confirmed by comparing it with the rest of the state of the art metaheuristics such as the AO, EO, PSO, WOA, and the original AHA algorithms. Table 1 presents a qualitative summary of the related literature by comparing existing thresholding and metaheuristic based segmentation methods in terms of key methodologies, objective criteria, qualitative metrics considered, strengths, limitations, and research gaps. This comparison highlights the limitations of conventional histogram based methods, standard metaheuristic algorithms, and initialization strategies, convergence stability, and population diversity. Table 1Qualitative summary of related literature and identified research gapsCategoryKey methodologiesObjective / criterionQualitative metricsconsideredStrengths, limitations, and researchgapsTraditionalthresholdingOtsu8, Kapur25Variance based and entropybased thresholdingHistogram distribution and gray level intensity informationThese methods are simple and widelyused for basic thresholding tasks.However, they mainly depend onhistogram information and do notinclude spatial contextual information,which may reduce stability in complexmultilevel segmentationStandardmetaheuristicsGA28, DE29, ABC30, PSO31, MFO32, AO33, EO34, WOA35Optimization of threshold values using entropy or threshold-based fitness functionsExploration ability, exploitation behaviour, convergence tendency, and threshold search performanceThese methods reduce thecomputational burden of exhaustivethreshold search. However, they maysuffer from premature convergenceand local optima stagnation in highdimensional search spaces, assuggested by the No Free LunchtheoremInitializationstrategiesLatin Hypercube5, Sobol, Halton,Sierpinski6Population diversity and search space coverageInitial solution distribution, population spread, and exploration stabilityStructured initialization improves initialpopulation diversity and reducesclustering of candidate solutions.Poorly structured initialization canlead to inefficient exploration andunstable convergence behaviour Table 1Qualitative summary of related literature and identified research gapsCategoryKey methodologiesObjective / criterionQualitative metricsconsideredStrengths, limitations, and researchgapsTraditionalthresholdingOtsu8, Kapur25Variance based and entropybased thresholdingHistogram distribution and gray level intensity informationThese methods are simple and widelyused for basic thresholding tasks.However, they mainly depend onhistogram information and do notinclude spatial contextual information,which may reduce stability in complexmultilevel segmentationStandardmetaheuristicsGA28, DE29, ABC30, PSO31, MFO32, AO33, EO34, WOA35Optimization of threshold values using entropy or threshold-based fitness functionsExploration ability, exploitation behaviour, convergence tendency, and threshold search performanceThese methods reduce thecomputational burden of exhaustivethreshold search. However, they maysuffer from premature convergenceand local optima stagnation in highdimensional search spaces, assuggested by the No Free LunchtheoremInitializationstrategiesLatin Hypercube5, Sobol, Halton,Sierpinski6Population diversity and search space coverageInitial solution distribution, population spread, and exploration stabilityStructured initialization improves initialpopulation diversity and reducesclustering of candidate solutions.Poorly structured initialization canlead to inefficient exploration andunstable convergence behaviour Qualitative summary of related literature and identified research gaps Traditional thresholding Variance based and entropy based thresholding Histogram distribution and gray level intensity information These methods are simple and widely used for basic thresholding tasks. However, they mainly depend on histogram information and do not include spatial contextual information, which may reduce stability in complex multilevel segmentation Standard metaheuristics GA28, DE29, ABC30, PSO31, MFO32, AO33, EO34, WOA35 Optimization of threshold values using entropy or threshold-based fitness functions Exploration ability, exploitation behaviour, convergence tendency, and threshold search performance These methods reduce the computational burden of exhaustive threshold search. However, they may suffer from premature convergence and local optima stagnation in high dimensional search spaces, as suggested by the No Free Lunch theorem Initialization strategies Latin Hypercube5, Sobol, Halton, Sierpinski6 Population diversity and search space coverage Initial solution distribution, population spread, and exploration stability Structured initialization improves initial population diversity and reduces clustering of candidate solutions. Poorly structured initialization can lead to inefficient exploration and unstable convergence behaviour Proposed work The proposed framework combines structured population initialization, energy curve based threshold evaluation, AHA search behaviour, and GDA based acceptance refinement for RGB multilevel image segmentation. The detailed mathematical steps are explained in the following subsections. Energy curve based multilevel thresholding Conventional histogram based methods do not effectively incorporate spatial contextual information when selecting suitable threshold values. The proposed method introduces an energy curve based multilevel thresholding strategy to address these limitations. The energy curve of an image is obtained via the modified Hopfield neural network9,10. The Hopfield neural network comprises several neurons, each interconnected by symmetric synaptic weights that exclude any self feedback loops. Hopfield formulated the energy function, which encapsulates the network’s total state by considering its entire architecture. In the Hopfield neural network, each pixel of the image is treated as a neuron, and its energy is computed. In addition to using pixel intensity distribution, it also uses adjacent interactions to capture local spatial patterns. The proposed method is applied to color images by performing multilevel segmentation on each RGB channel separately, in contrast to grayscale thresholding algorithms. A comparison of the original images, their corresponding histograms, and energy curves for six test samples is illustrated in Fig. 1. Algorithm 1Energy curve computation. Energy curve computation. Let the input images I be a color image composed of three channels \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I=\{{{I}_{R}},{{I}_{G}},{{I}_{B}}\}$$\end{document}, where each channel \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{I}^{c}}$$\end{document} (for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c\in \{R,G,B\}$$\end{document} is treated as a grayscale image of size m× n. For each channel, let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$g_{ij}^{c}$$\end{document} represents at pixel location (i, j) . For a given spatial position (i, j), the neighborhood system N of order X is defined as follows: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N_{ij}^{X}=(i+u,j+v)$$\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}$$(u,v)\in {{N}^{X}}$$\end{document}10. This defines the spatial correlation between nearby pixels of the image I. The configuration of the neighborhood system N varies based on order X9. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(u,v)\in \{(\pm 1,0),(0,\pm 1),(1,\pm 1),(-1,\pm 1)\}$$\end{document} is that of the second order neighborhood system \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N_{ij}^{X}$$\end{document}. In Table 2, the pixel at (i, j) spatial location is represented by its second order neighbor pixels. This 8 connectivity system enables modeling of local pixel interaction for energy evaluation. The neighborhood size acts as a tunable parameter dictating the algorithm’s sensitivity to local intensity shifts. Increasing the neighbourhood window can capture wider spatial information, but it also increases the computational cost. Very large windows are not suitable for precise thresholding because they may over-smooth the energy curve and blur fine regional boundaries. A tightly localized neighborhood is widely adopted in current segmentation literature because it strikes the necessary balance, preserving structural details while maintaining computational efficiency. Algorithm 1 summarizes the energy curve computation for RGB images. Table 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{N}^{2}}$$\end{document} neighborhood of pixel (i, j). \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(i\!-\!1,\ j\!-\!1)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(i\!-\!1,\ j)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(i\!-\!1,\ j\!+\!1)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(i,\ j\!-\!1)$$\end{document} (i,\ j) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(i,\ j\!+\!1)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(i\!+\!1,\ j\!-\!1)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(i\!+\!1,\ j)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(i\!+\!1,\ j\!+\!1)$$\end{document} Table 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{N}^{2}}$$\end{document} neighborhood of pixel (i, j). \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(i\!-\!1,\ j\!-\!1)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(i\!-\!1,\ j)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(i\!-\!1,\ j\!+\!1)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(i,\ j\!-\!1)$$\end{document} (i,\ j) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(i,\ j\!+\!1)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(i\!+\!1,\ j\!-\!1)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(i\!+\!1,\ j)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(i\!+\!1,\ j\!+\!1)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{N}^{2}}$$\end{document} neighborhood of pixel (i, j). For each gray level l∈ [0,L] , L is the maximum gray value i.e. 255, a binary matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$B_{l}^{c}=\{b_{ij}^{c}\}$$\end{document} is constructed in Eq. (1) for channel c such that:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} b_{ij}^{c} = {\left\{ \begin{array}{ll} 1, & \text {if } b_{ij}^{c} \le l \\ -1, & \text {otherwise} \end{array}\right. } \end{aligned}$$\end{document}In the binary matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$B_{l}$$\end{document}, every element \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b_{ij}$$\end{document} is assigned a value of either 1 or -1, based on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b_{ij}$$\end{document} of the corresponding pixel and the value of b. Consider a constant matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Z = \{z_{ij} \mid 1 \le i \le m,\, 1 \le j \le n\}$$\end{document} such that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$z_{ij} = 1$$\end{document} for all spatial positions (i, j). For the image I, the energy level l is computed as2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {{E}^{f}}_{l}=-\sum \limits _{i=1}^{m}{\sum \limits _{j=1}^{n}{\sum \limits _{N_{ij}^{2}}{{{b}_{ij}}}}}{{b}_{pq}}+\sum \limits _{i=1}^{m}{\sum \limits _{j=1}^{n}{\sum \limits _{N_{ij}^{2}}{{{z}_{ij}}}}}{{z}_{pq}} \end{aligned}$$\end{document}where (p, q) are coordinates of neighboring pixels of the pixel at (i, j), f=1 for grayscale image and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f=\{1,2,3\}$$\end{document} for a color image. To transition from the single-level energy evaluation in Eq. (3) to the complete 1D energy curve used for thresholding, the total spatial energy \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${E^{f}}_{l}$$\end{document} must be calculated for every possible gray level l ∈ [0, L]. In this way, the distinct energy states are grouped together to form a single continuous 1D landscape that can be described mathematically using the vector:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{E}^{f} = [ {E^{f}}_{0}, {E^{f}}_{1}, \dots , {E^{f}}_{L} ] \end{aligned}$$\end{document}Where the components are the spatial energy states of the entire image, taking into account the current threshold. Using this single dimension array, the vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{E}^{f}$$\end{document} replaces the typical one-dimensional intensity histogram, which is spatial in nature. Fig. 1Comparison of original images, histograms, and energy curves for six test samples from the BSDS500 and NASA datasets. Comparison of original images, histograms, and energy curves for six test samples from the BSDS500 and NASA datasets. Minimum cross entropy for multilevel thresholding In multilevel thresholding, the objective is to find threshold values that divide the image into meaningful regions. In the proposed method, MCEM is used as the objective function because it minimizes the difference between the original and segmented images. The related mathematical formulation is given below. Cross entropy was first presented by Kullback44 in 1968 under the name of directed divergence. Information theory is measured using cross entropy. The cross entropy between the two probability distributions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A=\{{{A}_{1}}, {{A}_{2}}, ...... {{A}_{n}}\} \&\text {B}= \{{{\text {B}}_{1}}, {{\text {B}}_{2}}, ...... {{\text {B}}_{n}}\}$$\end{document} is defined in Eq. (4):4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} c(A,B)=\sum \limits _{x=0}^{l-1}{{{A}_{x}}\log \frac{{{A}_{x}}}{{{B}_{x}}}} \end{aligned}$$\end{document}Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{A}_{x}}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{B}_{x}}$$\end{document} denote the pixel locations in the reference image and the segmented image, respectively. This can be regarded as an adaption of the maximum entropy method, wherein all \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{B}_{x}}$$\end{document} are originally allocated identical estimates. MCEM thresholding identifies the optimal threshold value by minimizing the cross entropy between the original image and its thresholded version. Let h(x) represent the histogram associated with the intensity levels x = 1, 2, … , l-1, where l denotes the total number of gray levels, and let I signify the original image. The thresholded image \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I_{TH}$$\end{document} was calculated using the threshold value (TH). The class wise mean intensity, as specified in Eq. (6), is calculated over the relevant histogram interval and is utilized in the formulation of reference probability distributions for MCEM.5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} I_{TH}(x,y) = {\left\{ \begin{array}{ll} \mu (1, TH), & \text {if } I(x,y) < TH \\ \mu (TH, l+1), & \text {if } I(x,y) \ge TH \end{array}\right. } \end{aligned}$$\end{document}Where,6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mu (u,v)=\sum \limits _{x=u}^{v-1}{xh(x)/}\sum \limits _{x=u}^{v-1}{h(x)} \end{aligned}$$\end{document}The cross entropy is computed by reformulating Eq. (4) as shown below to derive the entropy value, which serves as the objective function (commonly referred to as the fitness function)44:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \begin{aligned} F_{\text {cross}}(TH) = \sum \limits _{x=0}^{TH-1} x h(x) \log \left( \frac{x}{\mu (1, TH)} \right) \\ + \sum \limits _{x=TH}^{l-1} x h(x) \log \left( \frac{x}{\mu (TH, l-1)} \right) \end{aligned} \end{aligned}$$\end{document}With the extension of Eq. (7) to a multilevel method, this objective function takes into account a threshold value for bilevel thresholding41. Subsequently, Eq. (7) is formulated, and its explicit expression is given in Eq. (8)44.8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \begin{aligned} F_{\text {cross}}(TH) = \sum \limits _{x=0}^{l-1} x h(x) \log (x) \\ - \sum \limits _{x=0}^{TH-1} x h(x) \log \left( \mu (1, TH) \right) \\ - \sum \limits _{x=TH}^{l-1} x h(x) \log \left( \mu (TH, l-1) \right) \end{aligned} \end{aligned}$$\end{document}Based on the vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$TH=[T{{H}_{1}},T{{H}_{2}},....T{{H}_{tt}}]$$\end{document}, which comprises tt distinct threshold values, Eq. (7) multilevel method is carried out through the next Eqs. (9) and (10):9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} F_{\text {cross}}(\textbf{TH}) = \sum _{x=0}^{l-1} x h(x) \log (x) - \sum _{i=1}^{t} h_i \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h_i$$\end{document} is defined as follows and tt is the total number of thresholds.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \begin{aligned} h_1&= \sum _{x=1}^{TH_1 - 1} x h(x) \log \left( \mu (1, TH_1) \right) \\ h_k&= \sum _{x=TH_k}^{TH_{k+1} - 1} x h(x) \log \left( \mu (TH_k, TH_{k+1}) \right) , \quad 1< k < tt \\ h_{tt}&= \sum _{x=TH_{tt}}^{l - 1} x h(x) \log \left( \mu (TH_{tt}, l - 1) \right) \end{aligned} \end{aligned}$$\end{document} Population initialization strategies for enhanced diversity Initialization of the population plays an important role in the convergence stability and exploration ability of metaheuristic algorithms. In the proposed method, Latin Hypercube, Halton, Sobol, and Sierpinski strategies are jointly used during initialization. The combined effect of this initialization technique helps improve the diversity of the population, along with preventing premature convergence of the population in the search space. The exact mathematics involved in these initialization techniques is described in the following sub-sections. The initial population is divided into four parts, where each part is generated using one of these techniques. Latin hypercube Mathematically, the K-th sample point for the i-th dimension, denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{i}^{K}$$\end{document}, is generated using the following formulation:11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} x_{i}^{K} = \frac{\pi _{i}(K) - 1 + U_{i}^{K}}{N} \end{aligned}$$\end{document}where N represents the total number of partitions (which corresponds to the sub-population size), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K \in \{1, 2, \dots , N\}$$\end{document}. The term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pi _{i}$$\end{document} denotes a random permutation function applied to the sequence of integers \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{1, 2, \dots , N\}$$\end{document} for the i-th dimension. This permutation ensures that the samples do not align diagonally but are stochastically scattered. Finally, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$U_{i}^{K}$$\end{document} represents an independently and uniformly distributed random variable sampled within the continuous interval [0, 1]. By incorporating the permutation function and the uniform random variable, LHS prevents premature spatial clustering and ensures comprehensive global exploration during the initial iterations. Sobol Sobol sequence is designed with the aim of minimizing discrepancy using bitwise operations thus giving well spaced points in high dimensional space, is represented by using Eq. (12)5.12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} S(i)=\underset{j=1}{\overset{M}{\mathop {\oplus }}}\,{{g}_{j}}.{{V}_{j}} \end{aligned}$$\end{document}where i is expressed in binary, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$g_{j}$$\end{document} are its bits, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V_{j}$$\end{document} are direction numbers, and ⊕ denotes bitwise XOR. This construction minimizes discrepancy in multidimensional sampling. Halton The Halton sequence is represented in Eq. (13) built with coprime bases and provides a deterministic method of uniformly scattered points on dimensions5. 13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {{H}_{b}}(i)=\sum \limits _{j=0}^{M}{\frac{{{\mathbb {R}}_{j}}}{{{\mathbb {C}}^{j+1}}}} \end{aligned}$$\end{document} Where, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i = \mathbb {R}_{0} + \mathbb {R}_{1}\mathbb {C}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$+ \mathbb {R}_{2}\mathbb {C}^{2} + \dots + \mathbb {R}_{M}\mathbb {C}^{M}$$\end{document} is the expansion of i in base \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {b}.$$\end{document} Sierpinski The Sierpinski initialization in Eq. (14) provides a fractal distribution, that leads to better coverage of the search space by the recursive division of the range into smaller sub areas6.14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {{Z}_{i}}=\frac{{{\mathbb {Q}}_{i}}}{{{2}^{m}}},i=1,2,....,N \end{aligned}$$\end{document}Where, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mathbb {Q}}_{i}}$$\end{document} is an integer chosen according to the Sierpinski pattern, and m controls the recursion depth. Together, these four initialization strategies provide a more diverse starting population for the proposed search process. Local refinement via Great Deluge Algorithm GDA is used in the proposed method as a local refinement and adaptive acceptance mechanism. It is often compared with SA, but instead of using temperature as the acceptance barrier, GDA uses an upper boundary known as the water level4. Better solutions are always accepted, while some non improving solutions may also be accepted if their objective value remains within the current water level limit. This controlled acceptance supports more flexible local refinement. Compared with SA, GDA is simpler in terms of parameter handling because it mainly depends on the water level setting and its decay behaviour45,46. In minimization problems, a candidate solution can be retained if its objective value is less than or equal to the current water level. As the search continues, the water level is updated gradually to control the acceptance of candidate solutions during local refinement47. This makes GDA useful for improving local search while still allowing limited exploration in the earlier iterations. We adopt an adaptive GDA for local refinement in the proposed method. The initial water level is set based on the best initial objective value and the variation in the initial population. The initial water level is defined as:15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} W{{L}_{0}}=f({{\chi }_{best,0}})+\tau .{{\gamma }_{f,0}} \end{aligned}$$\end{document}Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\gamma }_{f,0}}$$\end{document} is the standard deviation, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\chi }_{best}}$$\end{document} denotes the best objective value among the initial solutions. The water level is then linearly decreased over the course of iterations, with a decay rate defined as:16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} decayRate=\frac{W{{L}_{0}}-{{\chi }_{best}}}{{{T}_{\max }}} \end{aligned}$$\end{document}Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{T}_{\max }}$$\end{document} is the maximum number of iterations. The water level is gradually reduced at each iteration as:17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} W{{L}_{t+1}}=W{{L}_{t}}-decayRate \end{aligned}$$\end{document}Candidate solutions χ ' is accepted when18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} f(\chi ')\le W{{L}_{t}}+{{\rho }_{t}} \end{aligned}$$\end{document}where the adaptive slack \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rho }_{t}}$$\end{document} decays over time and scales with the local quality gap.19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {{\rho }_{t}}={{\rho }_{0}}\exp (-\beta t)(\chi _{best}^{t}-\overline{{{\chi }^{t}}}) \end{aligned}$$\end{document}Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rho }_{0}}$$\end{document} is the initial slack β >0 is a decay factor, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\chi _{best}^{t}$$\end{document} is the best solution at iteration t, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{{{\chi }^{t}}}$$\end{document} is the mean objective of the current population. We determined the adaptive slack parameters namely the initial slack \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho _0$$\end{document} and the decay factor β through empirical trials across our benchmark images. This tuning was necessary to smoothly shift the algorithm from global exploration to local exploitation. The value of β determines how fast the constraints become tighter. However, the overall role of the adaptive slack is to avoid getting trapped in local minima for the problem. As the value of the slack is defined by the local gap between the best and mean solution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(\chi _{best}^{t}-\overline{\chi ^{t}})$$\end{document}, the mechanism automatically loosens the requirements if the individuals disperse or approach a local minimum. For a short time during which the algorithm becomes more lenient, the GDA can accept suboptimal solutions and acquire the motivation to get out of the local minima before the strict requirements come into play during the last steps. The generation of new individuals happens according to a diminishing step size:20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \chi '=\chi +{{s}_{t}}.\eta , {{s}_{t}}={{s}_{0}}{{(1-\frac{t}{{T}_{\max }})}^{\lambda }} \end{aligned}$$\end{document}Where η is a random perturbation vector, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{s}_{0}}$$\end{document} is the initial step size, λ >0 controls the decay profile. The algorithm iterates until a maximum number of iterations is reached or no significant improvement is observed, ultimately returning the best solution found. It should be noted that the GDA acceptance scheme is performed for the entire population during each cycle of the algorithm, but the actual evaluation of this scheme works on a per agent basis. In particular, for each new solution created by any hummingbird within the process of exploration of either the guided foraging behavior or the territorial one, this solution will be evaluated with respect to the current value of the water level in GDA. Fig. 2Foraging behaviors of the Artificial Hummingbird Algorithm. Foraging behaviors of the Artificial Hummingbird Algorithm. Proposed artificial hummingbird based energy curve driven multilevel thresholding In this section a bio inspired optimization algorithm, AHA, built on intelligent behaviors of hummingbirds is introduced3. The proposed method is utilized as a resilient and biologically inspired optimizer to tackle the multilevel threshold issue. The initially proposed technique draws inspiration from the intelligent foraging behavior of hummingbirds, their remarkable memory, and their diverse flight approaches incorporating the three primary components of AHA3. Biological modeling and core components of AHA The AHA simulates the foraging behaviour of hummingbirds and organizes the optimization process using three main components: food sources, hummingbirds, and the visit table. Food sources In AHA, each food source represents a candidate solution. For multilevel image thresholding, this solution is a set of threshold values used to divide the image intensity levels. The nectar refill rate of a food source is represented by its fitness value, where a better fitness value indicates a better threshold solution. Hummingbirds Each hummingbird is linked with a food source and uses memory to guide its movement. It can remember the quality of food sources and the time elapsed since each source was visited. This helps the search avoid repeated movement around the same region. Visit table The visit table records the visit status of each food source for every hummingbird. Sources that have not been visited for a longer time are given more attention, and when several sources have the same visit level, the one with better fitness is selected. The visit table is updated at every iteration to maintain a balance between exploration and exploitation. The AHA is used as the main swarm based optimizer for multilevel threshold selection. Based on the foraging behaviour and flight patterns of hummingbirds shown in Fig. 2, AHA applies three search strategies: guided foraging, territorial foraging, and migration foraging. These strategies help coordinate global exploration and local exploitation during the segmentation process. Fig. 3Three hummingbird flight behaviors: (a) Axial flight, (b) Diagonal flight, and (c) Omnidirectional flight. Three hummingbird flight behaviors: (a) Axial flight, (b) Diagonal flight, and (c) Omnidirectional flight. Guided foraging In guided foraging, a hummingbird selects a target food source by considering both the visit history and the nectar quality of available sources. In the optimization process, this means that the agent moves toward a promising solution while also considering sources that have not been visited recently. This behaviour helps the algorithm explore useful regions of the search space without repeatedly focusing on the same solution. AHA uses a directional switch vector to three different flight patterns: axial, diagonal, and omnidirectional flights, as shown in Fig. 3. These flight patterns control the movement direction of the hummingbird in the D-dimensional search space. Axial flight allows movement along one dimension, diagonal flight allows movement along selected dimensions, and omnidirectional flight allows movement along all dimensions. The Eq. (21) defines the axial flight.21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} d^{i} = {\left\{ \begin{array}{ll} 1, & \text {if } i = \texttt {randi}([1,D]) \\ 0, & \text {otherwise} \end{array}\right. } \quad \text {for } i = 1, \ldots , D \end{aligned}$$\end{document}The diagonal flight is defined in Eq. (22).22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} d^{i} = {\left\{ \begin{array}{ll} 1, & \text {if } i = p(j), j \in [1, k], \\ & \quad p = \texttt {randperm}(k), k \in \left[ 2, \big \lceil r_1 \cdot (D-2) \big \rceil + 1 \right] \\ 0, & \text {otherwise} \end{array}\right. } \end{aligned}$$\end{document}for i = 1, … , D. According to Eq. (23), omnidirectional flight is defined.23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} d^{i} = 1, \quad \text {for } i = 1, \ldots , D \end{aligned}$$\end{document}Using these flight patterns, the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i^{th}$$\end{document} hummingbird generates a candidate solution around the selected target food source. The updated candidate position is calculated as:24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} & {{U}_{i}}(t+1)={{x}_{i,target}}(t)+A.d.({{x}_{i}}(t)-{{x}_{i,target}}(t)) \end{aligned}$$\end{document}25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} & A \sim \mathscr {N}(0,1) \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_i(t)$$\end{document} is the current position of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i^{th}$$\end{document} hummingbird at iteration t, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{i,target}(t)$$\end{document} is the selected target food source, A is a normally distributed random variable with mean 0 and standard deviation 1, and d controls the active movement direction. The generated candidate solution is compared with the current solution using the objective function. The position update is defined as:26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {{x}_{i}}(t+1)=\left\{ \begin{array}{ll} {{x}_{i}}(t), & \text {if } f({{x}_{i}}(t)) \le f({{U}_{i}}(t+1)) \\ {{U}_{i}}(t+1), & \text {if } f({{x}_{i}}(t)) > f({{U}_{i}}(t+1)) \end{array} \right. \end{aligned}$$\end{document}where f(.) denotes the objective function to be minimized. Thus, the candidate solution is accepted only when it improves the current fitness value. Visit table management and target selection The visit table guides the foraging activity of hummingbirds by recording the visit status of each food source. A higher visit value indicates that a food source has not been visited for a longer time, making it more attractive for selection. At each iteration, each hummingbird first considers the food sources with the highest visit level. If more than one source has the same visit level, the source with the better nectar refilling rate, represented by the minimum objective function value, is selected as the target for guided foraging as given in Eq. (25). After guided foraging, the visit levels of the accessible food sources are increased by one, while the selected target source is reset to zero. If the newly generated candidate solution does not improve the fitness value, the hummingbird retains its current position. However, if a better candidate solution is obtained, it replaces the previous food source, and the corresponding visit table entries are updated. In this case, the new food source will get more visitation preference among the other hummingbirds visitation table to encourage exploration and maintain diversity among the population. Figure 4 shows the visitation table with six food sources where each food source has only one hummingbird at the moment. Each cell in the table represents visitation values; this means that the value represents the number of time intervals passed after a particular food source was visited by the respective hummingbird. For example, in the Fig. 4 the cell with the blue border marked with 9 implies that hummingbird 2 has not visited food source 6 for nine time intervals. The six hummingbirds image shown in Fig. 4, is also used as a test image. This image consists of complicated texture, intensity variation, wing and feather details. Fig. 4Foraging visit table of six hummingbirds. Foraging visit table of six hummingbirds. Territorial foraging In territorial foraging, a hummingbird prefers to find a new food source once the nectar in its current flower is depleted. It avoids visiting food sources that are already being targeted by other hummingbirds. As a result, a hummingbird may easily relocate to a nearby area inside its own territory, where it might discover a new food supply that is superior to its present one. The following is the mathematical formula that represents the local search for a potential food source by hummingbirds in the territorial foraging strategy:27\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} & {{U}_{i}}(t+1)={{x}_{i}}(t)+B.d.{{x}_{i}}(t) \end{aligned}$$\end{document}28\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} & B \sim \mathscr {N}(0, 1) \end{aligned}$$\end{document}where the normal distribution is applicable to the territorial variable B. Exploiting its distinctive aerial capabilities, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {N} (0,1)$$\end{document} in Eq. (28), characterized by a mean of 0 and a standard deviation of 1, enables each hummingbird to identify a novel food source in nearby to its current location. The visit table must be revised subsequent to the execution of the territorial foraging strategy. Migration foraging In AHA, migration foraging represents the ability of a hummingbird to leave a poor food source and move to a more promising region. This behaviour improves global search in multilevel thresholding and helps the algorithm avoid repeated search around weak solutions. When the migration condition is satisfied, the hummingbird associated with the worst food source is relocated to a new position generated within the search space. The visit table is then updated accordingly. The new position is calculated as:29\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {x}_{\text {worst}}(t+1) = LOW + r \cdot (UP - LOW) \end{aligned}$$\end{document}where r is a random vector in [0,1], and LOW and UP represent the lower and upper bounds of the search space. Migration helps the algorithm to steer clear of areas that do not provide any gain in terms of the fitness function during either guided or territorial foraging. When a source of food is left unchanged for an extended period of time, the hummingbird associated with that area is motivated to search in a different region. In this way, diversification is fostered, and the danger of getting stuck in local minima or early convergence is reduced. At the same time, in case a good food source has been found, the other birds will be motivated to migrate to that region. Migration takes place after a number of iterations according to Eq. (30). A flowchart of the whole process is provided in Fig. 5.30\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} M=2n \end{aligned}$$\end{document} Fig. 5Operational framework of the proposed segmentation method. Operational framework of the proposed segmentation method. Computational complexity analysis The total computational complexity of the proposed hybrid framework is evaluated by analyzing its three primary stages. Let m × n represent the image dimensions, L be the maximum gray level (256), C be the number of color channels, D be the number of threshold dimensions, nPop be the population size, and T be the maximum number of iterations. Energy curve computation (Preprocessing): As detailed in Algorithm 1, this phase evaluates spatial neighborhood interactions for every pixel. For a neighborhood of size W, computing the energy for all gray levels across C channels requires \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(C \cdot L \cdot m \cdot n \cdot W)$$\end{document} operations. Since C and W are small constants, this mathematically simplifies to an asymptotic complexity of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(L \cdot m \cdot n)$$\end{document}. Initialization phase: Generating the initial population utilizing four structured sequences (Latin Hypercube, Sobol, Halton, and Sierpinski) requires assigning D dimensional variables for nPop individuals. The complexity for this step is strictly O(nPop · D). Optimization and refinement (AHA-GDA): During the AHA and GDA phase, the algorithm operates over T iterations. In each iteration, updating the positions (guided, territorial, or migration foraging) takes O(D) operations per individual. Furthermore, evaluating the MCEM fitness function for each updated candidate solution requires iterating through the histogram bins, taking O(L) time. Therefore, the cost per iteration for the entire population is O(nPop · (D + L)). Over T iterations, this results in an optimization complexity of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(T \cdot nPop \cdot (D + L))$$\end{document}. Total complexity: By summing the independent phases, the overall mathematical formulation of the algorithm’s time complexity is:31\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} O(L \cdot m \cdot n) + O(nPop \cdot D) + O(T \cdot nPop \cdot (D + L)) \end{aligned}$$\end{document}Because initialization O(nPop · D) is asymptotically smaller than the iterative optimization step, the final dominant computational complexity simplifies to:32\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} O(L \cdot m \cdot n + T \cdot nPop \cdot (D + L)) \end{aligned}$$\end{document} Mathematical formulation of component synergy and architecture comparison The integration of multiple hybrid components and prove their mathematical synergy, the state transition of the proposed algorithm can be modeled as a composite function. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {S}$$\end{document} represent the search space of possible threshold vectors. The complete synergistic progression of a population state P from iteration t to t+1 is defined by the following mapping:33\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} P_{t+1} = \mathscr {G} \left( \mathscr {A}(P_t), WL_t, \textbf{E}^f \right) \end{aligned}$$\end{document}This composite formulation mathematically binds the four isolated components: Initialization: At t=0, the structured sequences (Latin Hypercube, Sobol, Halton, Sierpinski) map an empty space to a highly uniformly distributed initial population:34\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} P_0 = \mathscr {I}(\mathscr {S}) \end{aligned}$$\end{document} AHA foraging (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {A}$$\end{document}): The AHA acts as the primary search operator, mapping the current population \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_t$$\end{document} to a set of new candidate solutions, denoted as:35\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} U_i = \mathscr {A}(x_i) \end{aligned}$$\end{document}Spatial energy evaluation (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{E}^f$$\end{document}): Unlike standard models that evaluate fitness f(x) using 1D histograms, the spatial energy curve \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{E}^f$$\end{document} evaluates the structural context of both the candidate solution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$U_i$$\end{document} and the current solution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_i$$\end{document}. GDA refinement (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathscr {G}$$\end{document}): The GDA acts as the final synergistic filter, judging the candidate solution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$U_i$$\end{document} against the dynamic water level \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$WL_t$$\end{document}. Comparison: standard AHA vs. proposed hybrid architecture The increased complexity of the proposed hybrid architecture is theoretically and empirically justified when compared to the standard AHA mechanism. In a standard AHA, the candidate acceptance is greedy. A generated candidate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$U_i$$\end{document} replaces the current solution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_i$$\end{document} for the next iteration only if it yields a better fitness:36\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} x_{i}(t+1) = {\left\{ \begin{array}{ll} U_i, & \text {if } f_{hist}(U_i) < f_{hist}(x_i) \\ x_i(t), & \text {otherwise} \end{array}\right. } \end{aligned}$$\end{document}While computationally simpler, this strict greediness combined with context blind 1D histograms (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f_{hist}$$\end{document}) causes standard AHA to frequently stagnate in local optima when processing highly textured images. Conversely, the proposed hybrid architecture utilizes spatially aware acceptance criterion governed by the GDA water level:37\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} x_{i}(t+1) = {\left\{ \begin{array}{ll} U_i, & \text {if } f_{EC}(U_i) < f_{EC}(x_i) \text { or } f_{EC}(U_i) \le WL_t \\ x_i(t), & \text {otherwise} \end{array}\right. } \end{aligned}$$\end{document}The algorithm allowing acceptance of low quality solutions is allowed temporarily if their value is still below the gradually decreasing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$WL_t$$\end{document}, which allows avoiding getting stuck in local optima. Results and discussion In addition to the suggested metrics for assessing segmentation outcomes, this section describes the computer specifications used for operations. Additionally, by using multilevel thresholding, it clarifies how metaheuristics are applied to solve the image segmentation problem. Experimental conditions The proposed method was evaluated using two public datasets: BSDS50011 and NASA12. The BSDS500 dataset contains 500 natural images, including objects, animals, buildings, vegetation, and outdoor scenes. These images have different textures, object boundaries, backgrounds, and illumination conditions. The NASA Earth Observatory dataset contains 44 satellite images. These images include land, vegetation, water bodies, clouds, deserts, and other earth surface regions. They contain dense spatial details and different colour variations, which make them suitable for testing satellite image segmentation. No preprocessing or annotation procedure was performed on the NASA dataset. The dataset images were directly utilized for experimental evaluation. Since the proposed method is not based on machine learning or deep learning techniques, no dataset splitting into training or testing subsets was required. The complete dataset was directly used for performance evaluation and comparison. All images taken from BSDS500 and NASA databases were used to perform the experiments. For the sake of brevity, the manuscript presents only representative segmented images in the result outputs, while the values reported in the result tables were obtained from the corresponding segmented images. The experiments were conducted using MATLAB R2023b software on the Windows 11 64-bit operating system equipped with an Intel Core i7 CPU and 16 GB of memory. All methods were tested using three threshold values: nT = 3, nT = 5, and nT = 8. To guarantee the comparability of the results, the same parameters were used for testing all the proposed approaches. The test consisted of 200 iterations with 30 individuals in the population. The color image was divided into the R, G, and B components. These components were processed separately during segmentation and then combined together to create the final segmented image. Code availability The MATLAB code for the proposed method is available in a public GitHub repository: https://github.com/suprajatirumalasetti/AHA_GDA_Image_Segmentation_Code. Quantitative evaluation Evaluating the quality of segmented images has become crucial as image segmentation techniques have expanded in recent years. Seven widely accepted image quality metrics were employed: PSNR13, SSIM14, FSIM15, QILV16, Correlation Coefficient17, EPI18, and MIF19. Peak signal to noise ratio The PSNR metric defines the extent of fidelity achieved by segmenting the image compared to the original image, which is measured in dB. Higher PSNR metrics imply lesser distortion and better retention of image information through the process of image segmentation. Table 3 provides the PSNR measures for BSDS500 and NASA images using thresholds of 3, 5, and 8. It can be seen that the proposed technique generally produces higher PSNR measurements as compared to the other comparative algorithms, indicating the effectiveness of the former in retaining image information through segmentation. This holds true for both types of images (BSDS500 and NASA) and implies that the proposed technique retains information more effectively during segmentation. This inference is substantiated by the fact that the proposed technique outperforms all other baseline techniques. Fig. 6Comparison of AO, EO, PSO, WOA, and proposed method from top to bottom for Red, Green, Blue channel thresholding and final segmentation results for the image 247003 of 8 level threshold from left to right. Comparison of AO, EO, PSO, WOA, and proposed method from top to bottom for Red, Green, Blue channel thresholding and final segmentation results for the image 247003 of 8 level threshold from left to right. Fig. 7Comparison of 120093 three level threshold with segmented and histogram images with AO, EO, PSO, WOA, and proposed method from top to bottom. Comparison of 120093 three level threshold with segmented and histogram images with AO, EO, PSO, WOA, and proposed method from top to bottom. Table 3Comparative PSNR results on BSDS500 and NASA datasets with varying threshold levels.BSDS500NASAImage namenTAOEOPSOWOAProposedImage nameTAHAAOEOPSOWOAProposed120093324.729324.574824.399924.5541 30.4735 Kizimen333.956731.725632.069133.335033.8267 33.9567 525.330724.958424.784125.3966 33.6325 536.322432.561033.973235.827733.7256 36.3300 825.444325.134924.996524.8316 34.6113 837.235529.738637.604438.397231.2038 38.4143 159002325.540425.634925.335425.6329 29.3680 Yellowstone333.183036.681833.976936.018734.0118 33.5834 526.063326.848925.688026.2479 30.1126 535.580431.549132.257733.394731.3954 35.7575 826.951226.960025.516425.9203 31.3913 840.135436.593630.309939.588240.3297 40.3536 189029325.506925.418525.238025.3996 28.9393 Goldpanfire335.101930.655234.746432.521030.9620 35.1292 525.895826.238625.457125.7588 29.5218 535.024232.125733.488932.070832.3470 36.2502 826.364126.580525.415926.6839 30.3501 839.272636.818533.916427.737131.1206 39.4828 228076325.365025.399425.232125.4777 30.7777 Nyamuragira331.542930.147630.280629.642230.3468 31.6743 526.213625.847125.502525.9352 32.5953 535.135034.838827.361831.918031.1153 35.2226 826.593526.406625.583326.2526 36.2204 836.878831.130227.260635.239327.1082 37.3473 247003325.110825.063924.604925.1223 30.2612 Rockhouse335.146530.711634.841831.910032.3914 35.1465 525.981025.717724.992625.4676 33.1653 537.595728.517436.322836.608035.5939 37.8557 826.079225.696824.694525.6616 35.9528 839.994433.762634.021039.362432.9135 40.3916 326085325.166825.000524.779825.0315 31.0340 Yarrabin333.012632.275331.474932.358730.9591 33.0856 525.503925.557524.937925.5555 33.2250 534.620632.050432.223134.559131.1201 34.6871 826.096025.793825.006925.6344 36.5538 839.929828.007832.449337.983225.9390 39.9216 Significant values are in [bold]. Table 3Comparative PSNR results on BSDS500 and NASA datasets with varying threshold levels.BSDS500NASAImage namenTAOEOPSOWOAProposedImage nameTAHAAOEOPSOWOAProposed120093324.729324.574824.399924.5541 30.4735 Kizimen333.956731.725632.069133.335033.8267 33.9567 525.330724.958424.784125.3966 33.6325 536.322432.561033.973235.827733.7256 36.3300 825.444325.134924.996524.8316 34.6113 837.235529.738637.604438.397231.2038 38.4143 159002325.540425.634925.335425.6329 29.3680 Yellowstone333.183036.681833.976936.018734.0118 33.5834 526.063326.848925.688026.2479 30.1126 535.580431.549132.257733.394731.3954 35.7575 826.951226.960025.516425.9203 31.3913 840.135436.593630.309939.588240.3297 40.3536 189029325.506925.418525.238025.3996 28.9393 Goldpanfire335.101930.655234.746432.521030.9620 35.1292 525.895826.238625.457125.7588 29.5218 535.024232.125733.488932.070832.3470 36.2502 826.364126.580525.415926.6839 30.3501 839.272636.818533.916427.737131.1206 39.4828 228076325.365025.399425.232125.4777 30.7777 Nyamuragira331.542930.147630.280629.642230.3468 31.6743 526.213625.847125.502525.9352 32.5953 535.135034.838827.361831.918031.1153 35.2226 826.593526.406625.583326.2526 36.2204 836.878831.130227.260635.239327.1082 37.3473 247003325.110825.063924.604925.1223 30.2612 Rockhouse335.146530.711634.841831.910032.3914 35.1465 525.981025.717724.992625.4676 33.1653 537.595728.517436.322836.608035.5939 37.8557 826.079225.696824.694525.6616 35.9528 839.994433.762634.021039.362432.9135 40.3916 326085325.166825.000524.779825.0315 31.0340 Yarrabin333.012632.275331.474932.358730.9591 33.0856 525.503925.557524.937925.5555 33.2250 534.620632.050432.223134.559131.1201 34.6871 826.096025.793825.006925.6344 36.5538 839.929828.007832.449337.983225.9390 39.9216 Significant values are in [bold]. Comparative PSNR results on BSDS500 and NASA datasets with varying threshold levels. Significant values are in [bold]. Table 4Comparative SSIM results on BSDS500 and NASA datasets at varying threshold levels.BSDS500NASAImage namenTAOEOPSOWOAProposedImage namenTAHAAOEOPSOWOAProposed12009330.96050.95950.8916 0.9627 0.9570Kizimen3 0.9491 0.92810.94230.89760.94100.949150.97970.97710.92740.9811 0.9864 50.97560.95950.95710.88660.9450 0.9763 80.98580.98050.92750.9671 0.9947 80.98920.98730.95470.87330.9614 0.9905 15900230.9679 0.9762 0.86740.97610.9748Yellowstone3 0.9561 0.88660.89470.78970.89350.956150.98390.98550.94210.9828 0.9897 50.9800 0.9845 0.91970.85400.93950.98038 0.9912 0.98660.91080.97940.994380.99010.98630.97510.84060.9244 0.9903 18902930.97100.97510.87060.9761 0.9816 goldpanfire30.95190.95080.93850.91590.9020 0.9530 50.98450.98650.87360.9874 0.9904 50.97830.95090.94400.83290.9528 0.9783 80.99120.98700.88350.9904 0.9953 80.98800.98120.95220.86370.9636 0.9888 22807630.9725 0.9776 0.90820.97100.9755Nyamuragira30.95430.93800.94270.87940.9471 0.9683 50.98770.98490.92040.9884 0.9910 5 0.9861 0.96540.97040.88020.96420.986180.9844 0.9901 0.96320.98480.993780.99150.99110.96360.88200.9773 0.9937 24700330.96700.97020.897 0.9725 0.9629Rockhouse3 0.9586 0.94040.95480.91050.9492 0.9586 50.98250.98140.93800.9833 0.9882 50.9763 0.9833 0.96670.92240.97010.976480.98950.98390.90680.9776 0.9950 80.98790.98490.97120.88300.9529 0.9879 32608530.97030.96700.90090.9682 0.9759 Yarrabin3 0.9693 0.96440.94410.91880.9452 0.9693 50.98030.98170.90370.9839 0.9891 50.98310.98070.95720.84960.9443 0.9834 80.98880.98490.93290.9847 0.9946 80.99140.98020.96500.89180.9656 0.9917 Significant values are in [bold]. Table 4Comparative SSIM results on BSDS500 and NASA datasets at varying threshold levels.BSDS500NASAImage namenTAOEOPSOWOAProposedImage namenTAHAAOEOPSOWOAProposed12009330.96050.95950.8916 0.9627 0.9570Kizimen3 0.9491 0.92810.94230.89760.94100.949150.97970.97710.92740.9811 0.9864 50.97560.95950.95710.88660.9450 0.9763 80.98580.98050.92750.9671 0.9947 80.98920.98730.95470.87330.9614 0.9905 15900230.9679 0.9762 0.86740.97610.9748Yellowstone3 0.9561 0.88660.89470.78970.89350.956150.98390.98550.94210.9828 0.9897 50.9800 0.9845 0.91970.85400.93950.98038 0.9912 0.98660.91080.97940.994380.99010.98630.97510.84060.9244 0.9903 18902930.97100.97510.87060.9761 0.9816 goldpanfire30.95190.95080.93850.91590.9020 0.9530 50.98450.98650.87360.9874 0.9904 50.97830.95090.94400.83290.9528 0.9783 80.99120.98700.88350.9904 0.9953 80.98800.98120.95220.86370.9636 0.9888 22807630.9725 0.9776 0.90820.97100.9755Nyamuragira30.95430.93800.94270.87940.9471 0.9683 50.98770.98490.92040.9884 0.9910 5 0.9861 0.96540.97040.88020.96420.986180.9844 0.9901 0.96320.98480.993780.99150.99110.96360.88200.9773 0.9937 24700330.96700.97020.897 0.9725 0.9629Rockhouse3 0.9586 0.94040.95480.91050.9492 0.9586 50.98250.98140.93800.9833 0.9882 50.9763 0.9833 0.96670.92240.97010.976480.98950.98390.90680.9776 0.9950 80.98790.98490.97120.88300.9529 0.9879 32608530.97030.96700.90090.9682 0.9759 Yarrabin3 0.9693 0.96440.94410.91880.9452 0.9693 50.98030.98170.90370.9839 0.9891 50.98310.98070.95720.84960.9443 0.9834 80.98880.98490.93290.9847 0.9946 80.99140.98020.96500.89180.9656 0.9917 Significant values are in [bold]. Comparative SSIM results on BSDS500 and NASA datasets at varying threshold levels. Significant values are in [bold]. Table 5FSIM comparison results for different algorithms on BSDS500 and NASA dataset images.BSDS500NASAImage namenTAOEOPSOWOAProposedImage nameT AHA AOEOPSOWOAProposed12009330.7917 0.7994 0.67920.79370.7583Kizimen3 0.8048 0.76100.79950.56800.78660.804850.82280.82480.71870.8231 0.8408 50.89980.84520.86080.57190.7925 0.9011 80.85470.83630.67580.8113 0.8971 80.91470.92650.84080.52980.8343 0.9544 15900230.6909 0.7184 0.58170.71890.7004Yellowstone30.80500.73170.75520.58290.7292 0.8060 50.77090.79030.70340.7597 0.8129 50.89680.88620.79170.61830.8457 0.8988 80.84280.78910.60150.7437 0.8936 80.94550.92890.89070.59490.7919 0.9537 18902930.73780.76020.6785 0.7637 0.7472Goldpanfire30.80340.77600.80050.70110.7081 0.8074 50.79670.80950.7025 0.8176 0.82025 0.9014 0.83500.80330.48210.84560.899280.85180.82290.68310.8388 0.8805 80.95430.92320.83000.57640.8556 0.9551 22807630.8299 0.8388 0.76240.81850.7935Nyamuragira30.69450.63790.63590.59980.6156 0.6955 50.85020.86740.8013 0.8777 0.86635 0.7951 0.71570.72170.53790.65920.794880.88200.87520.78820.8502 0.9053 80.88210.84100.73850.59770.7257 0.8824 24700330.69110.69010.5581 0.6939 0.6406Rockhouse30.75120.70910.74680.55750.7294 0.7650 50.74010.74550.62300.7479 0.7651 5 0.8760 0.86300.81870.56200.82550.875680.80800.75440.56240.7182 0.8599 80.93100.90020.81940.46280.7163 0.9406 32608530.71140.69670.58080.7072 0.7105 Yarrabin30.71200.71280.70150.56930.6877 0.7320 50.76920.79730.62720.8115 0.8355 5 0.8556 0.81940.76070.44250.70590.853980.84770.83190.63080.8045 0.9020 80.93260.86560.80010.45580.8157 0.9334 Significant values are in [bold]. Table 5FSIM comparison results for different algorithms on BSDS500 and NASA dataset images.BSDS500NASAImage namenTAOEOPSOWOAProposedImage nameT AHA AOEOPSOWOAProposed12009330.7917 0.7994 0.67920.79370.7583Kizimen3 0.8048 0.76100.79950.56800.78660.804850.82280.82480.71870.8231 0.8408 50.89980.84520.86080.57190.7925 0.9011 80.85470.83630.67580.8113 0.8971 80.91470.92650.84080.52980.8343 0.9544 15900230.6909 0.7184 0.58170.71890.7004Yellowstone30.80500.73170.75520.58290.7292 0.8060 50.77090.79030.70340.7597 0.8129 50.89680.88620.79170.61830.8457 0.8988 80.84280.78910.60150.7437 0.8936 80.94550.92890.89070.59490.7919 0.9537 18902930.73780.76020.6785 0.7637 0.7472Goldpanfire30.80340.77600.80050.70110.7081 0.8074 50.79670.80950.7025 0.8176 0.82025 0.9014 0.83500.80330.48210.84560.899280.85180.82290.68310.8388 0.8805 80.95430.92320.83000.57640.8556 0.9551 22807630.8299 0.8388 0.76240.81850.7935Nyamuragira30.69450.63790.63590.59980.6156 0.6955 50.85020.86740.8013 0.8777 0.86635 0.7951 0.71570.72170.53790.65920.794880.88200.87520.78820.8502 0.9053 80.88210.84100.73850.59770.7257 0.8824 24700330.69110.69010.5581 0.6939 0.6406Rockhouse30.75120.70910.74680.55750.7294 0.7650 50.74010.74550.62300.7479 0.7651 5 0.8760 0.86300.81870.56200.82550.875680.80800.75440.56240.7182 0.8599 80.93100.90020.81940.46280.7163 0.9406 32608530.71140.69670.58080.7072 0.7105 Yarrabin30.71200.71280.70150.56930.6877 0.7320 50.76920.79730.62720.8115 0.8355 5 0.8556 0.81940.76070.44250.70590.853980.84770.83190.63080.8045 0.9020 80.93260.86560.80010.45580.8157 0.9334 Significant values are in [bold]. FSIM comparison results for different algorithms on BSDS500 and NASA dataset images. Significant values are in [bold]. Table 6QILV comparison results for different algorithms on BSDS500 and NASA dataset images.BSDS500NASAImage namenTAOEOPSOWOAProposedImage namenTAHAAOEOPSOWOAProposed12009330.9141 0.9187 0.57820.91410.9094Kizimen3 0.8456 0.83040.83020.80200.83670.845650.95520.95540.46660.9166 0.9610 50.91700.88200.88050.81150.8407 0.9196 80.98430.95890.79260.9429 0.9853 80.96200.95800.86600.80810.9079 0.9677 15900230.89020.89230.60880.8919 0.9270 Yellowstone30.81380.67880.69170.58120.6722 0.8149 50.94790.93620.65210.9175 0.9768 50.90510.92660.73300.72540.7944 0.9063 80.97530.94140.72460.9310 0.9834 80.95730.94410.89860.67630.7627 0.9592 18902930.93270.92370.61330.9248 0.9686 Goldpanfire30.8327 0.8473 0.81550.74700.71880.832750.94140.94720.52990.9502 0.9815 50.91210.84830.80200.72730.8459 0.9136 80.96920.96950.44180.9362 0.9930 80.95580.94820.82950.56680.8794 0.9567 22807630.82350.91330.33860.9111 0.9231 Nyamuragira30.91150.85130.85800.77240.8683 0.9125 50.95370.94930.41640.9376 0.9711 50.95790.92240.94590.80670.9096 0.9583 80.95290.95480.57550.9249 0.9758 80.9828 0.9831 0.92260.86410.95710.98132470033 0.9202 0.91220.81650.88540.8990Rockhouse30.86680.82140.85990.81860.8324 0.8688 5 0.9607 0.93970.86600.93820.963650.9199 0.9607 0.90040.85850.90180.920680.97450.93990.77870.9497 0.9836 80.95580.94510.89680.60260.8598 0.9567 32608530.90080.89910.64940.8988 0.9678 Yarrabin30.90030.88780.82270.82530.8118 0.9009 50.94310.94140.77340.9402 0.9800 50.9390 0.9508 0.85070.82780.81370.940180.97970.94120.15830.9664 0.9874 80.97060.92870.87330.82460.9002 0.9736 Significant values are in [bold]. Table 6QILV comparison results for different algorithms on BSDS500 and NASA dataset images.BSDS500NASAImage namenTAOEOPSOWOAProposedImage namenTAHAAOEOPSOWOAProposed12009330.9141 0.9187 0.57820.91410.9094Kizimen3 0.8456 0.83040.83020.80200.83670.845650.95520.95540.46660.9166 0.9610 50.91700.88200.88050.81150.8407 0.9196 80.98430.95890.79260.9429 0.9853 80.96200.95800.86600.80810.9079 0.9677 15900230.89020.89230.60880.8919 0.9270 Yellowstone30.81380.67880.69170.58120.6722 0.8149 50.94790.93620.65210.9175 0.9768 50.90510.92660.73300.72540.7944 0.9063 80.97530.94140.72460.9310 0.9834 80.95730.94410.89860.67630.7627 0.9592 18902930.93270.92370.61330.9248 0.9686 Goldpanfire30.8327 0.8473 0.81550.74700.71880.832750.94140.94720.52990.9502 0.9815 50.91210.84830.80200.72730.8459 0.9136 80.96920.96950.44180.9362 0.9930 80.95580.94820.82950.56680.8794 0.9567 22807630.82350.91330.33860.9111 0.9231 Nyamuragira30.91150.85130.85800.77240.8683 0.9125 50.95370.94930.41640.9376 0.9711 50.95790.92240.94590.80670.9096 0.9583 80.95290.95480.57550.9249 0.9758 80.9828 0.9831 0.92260.86410.95710.98132470033 0.9202 0.91220.81650.88540.8990Rockhouse30.86680.82140.85990.81860.8324 0.8688 5 0.9607 0.93970.86600.93820.963650.9199 0.9607 0.90040.85850.90180.920680.97450.93990.77870.9497 0.9836 80.95580.94510.89680.60260.8598 0.9567 32608530.90080.89910.64940.8988 0.9678 Yarrabin30.90030.88780.82270.82530.8118 0.9009 50.94310.94140.77340.9402 0.9800 50.9390 0.9508 0.85070.82780.81370.940180.97970.94120.15830.9664 0.9874 80.97060.92870.87330.82460.9002 0.9736 Significant values are in [bold]. QILV comparison results for different algorithms on BSDS500 and NASA dataset images. Significant values are in [bold]. Fig. 8Comparison of AO, EO, PSO, WOA, and proposed method from top to bottom for Red, Green, Blue channel thresholding and final segmentation results for the image Yellowstone of 3 level threshold from left to right. Fig. 9Comparison of Goldpanfire five level threshold with segmented and histogram images with AO, EO, PSO, WOA and proposed from top to bottom, and Proposed method with segmented images and energy curve. Comparison of AO, EO, PSO, WOA, and proposed method from top to bottom for Red, Green, Blue channel thresholding and final segmentation results for the image Yellowstone of 3 level threshold from left to right. Comparison of Goldpanfire five level threshold with segmented and histogram images with AO, EO, PSO, WOA and proposed from top to bottom, and Proposed method with segmented images and energy curve. Structural similarity index measure SSIM was used to evaluate how well the segmented image preserves the structural information of the original image. Which takes into account not only pixel-to-pixel comparison but also considers luminance, contrast, and structure of the compared images. Therefore, SSIM can be regarded as the metric for verifying whether the segmentation algorithm is able to preserve structure of the initial image. According to Table 4, the values obtained by the proposed segmentation algorithm for both BSDS500 and NASA datasets are quite consistent regardless of the threshold value used. This suggests that the proposed method preserves structural information in both natural and satellite images with various boundary and texture features. Feature similarity index measure FSIM was utilized as a measure for feature level analysis to identify whether segmentation has retained any key visual features from the original image. Unlike pixel level analysis, FSIM uses perceptually significant information in its calculation, including phase congruency and gradient magnitude. Hence, FSIM is better suited for analyzing the quality of edge, texture, and local feature preservation by multilevel thresholding. The FSIM results reported in Table 5 provide a feature-level insight into the quality of segmentation results. As can be seen from the table, the new technique produces consistent results irrespective of the threshold level used and does not experience as steep a drop compared to some other baseline techniques. Therefore, we may conclude that the introduced approach is more efficient at preserving local visual features especially in textured images. Quality index based on local variance The local variance acts as the index that will determine how well small variations and local intensities have been retained after segmentation. Therefore, in this research, the QILV has been used as the evaluation metric. The higher the value of QILV, the better the ability of the segmented image to retain local variations. In multilevel thresholding, the loss of local variations is a major problem. Therefore, the higher value of QILV shows the strength of this approach to preserve local variations even with the use of high thresholds. As shown in Table 6, the proposed method has performed better as compared to the other methods. This performance increases with the increase in the threshold while other methods show deterioration in performance. Table 7Correlation comparison results for different algorithms on BSDS500 and NASA dataset images.BSDS500NASAImage namenTAOEOPSOWOAProposedImage namenTAHAAOEOPSOWOAProposed12009330.71460.73910.41350.7255 0.7497 Kizimen30.56110.48590.55750.25630.5250 0.5613 50.76570.74930.35000.7423 0.8335 50.69490.56090.64080.29520.5340 0.6929 80.85960.78350.43530.7598 0.8858 80.76970.69930.59330.29130.5322 0.7727 15900230.47990.58720.34010.5961 0.6310 Yellowstone30.63040.46800.58380.29380.5868 0.6287 50.57130.65310.31650.5805 0.7460 50.73830.65520.65450.32360.6861 0.7401 80.77620.70270.38810.6404 0.8299 80.82610.76850.66020.38000.5862 0.8271 18902930.55970.57800.43090.5800 0.5960 Goldpanfire30.58540.43790.51410.44570.4822 0.5854 50.62960.62870.36900.5780 0.7843 50.68120.58950.57970.19670.6043 0.6816 80.63760.62310.35380.6611 0.8503 80.78680.69120.54770.21820.5729 0.7878 22807630.59200.68620.40310.6809 0.7069 Nyamuragira30.30120.25720.26800.22600.2577 0.3061 50.75040.76490.39260.7489 0.8191 50.43270.29270.30270.18870.2752 0.4303 80.75910.77190.42610.7203 0.8745 80.56090.44280.31730.19190.3117 0.5704 24700330.44920.48610.30020.4391 0.5276 Rockhouse30.41290.36770.36030.25140.3974 0.4288 50.51690.58090.31980.5189 0.6619 50.67120.50840.53780.25390.5579 0.6672 80.65100.58540.32080.5441 0.7809 80.74100.56040.48760.14350.3655 0.7410 32608530.38450.52360.39270.5175 0.5861 Yarrabin30.33800.28170.31330.23290.3435 0.3380 50.58870.58440.34530.5916 0.7029 50.55030.27600.31880.13830.3015 0.5526 80.67460.63280.08730.5536 0.7957 80.68190.47370.42500.14530.4233 0.6851 Significant values are in [bold]. Table 7Correlation comparison results for different algorithms on BSDS500 and NASA dataset images.BSDS500NASAImage namenTAOEOPSOWOAProposedImage namenTAHAAOEOPSOWOAProposed12009330.71460.73910.41350.7255 0.7497 Kizimen30.56110.48590.55750.25630.5250 0.5613 50.76570.74930.35000.7423 0.8335 50.69490.56090.64080.29520.5340 0.6929 80.85960.78350.43530.7598 0.8858 80.76970.69930.59330.29130.5322 0.7727 15900230.47990.58720.34010.5961 0.6310 Yellowstone30.63040.46800.58380.29380.5868 0.6287 50.57130.65310.31650.5805 0.7460 50.73830.65520.65450.32360.6861 0.7401 80.77620.70270.38810.6404 0.8299 80.82610.76850.66020.38000.5862 0.8271 18902930.55970.57800.43090.5800 0.5960 Goldpanfire30.58540.43790.51410.44570.4822 0.5854 50.62960.62870.36900.5780 0.7843 50.68120.58950.57970.19670.6043 0.6816 80.63760.62310.35380.6611 0.8503 80.78680.69120.54770.21820.5729 0.7878 22807630.59200.68620.40310.6809 0.7069 Nyamuragira30.30120.25720.26800.22600.2577 0.3061 50.75040.76490.39260.7489 0.8191 50.43270.29270.30270.18870.2752 0.4303 80.75910.77190.42610.7203 0.8745 80.56090.44280.31730.19190.3117 0.5704 24700330.44920.48610.30020.4391 0.5276 Rockhouse30.41290.36770.36030.25140.3974 0.4288 50.51690.58090.31980.5189 0.6619 50.67120.50840.53780.25390.5579 0.6672 80.65100.58540.32080.5441 0.7809 80.74100.56040.48760.14350.3655 0.7410 32608530.38450.52360.39270.5175 0.5861 Yarrabin30.33800.28170.31330.23290.3435 0.3380 50.58870.58440.34530.5916 0.7029 50.55030.27600.31880.13830.3015 0.5526 80.67460.63280.08730.5536 0.7957 80.68190.47370.42500.14530.4233 0.6851 Significant values are in [bold]. Correlation comparison results for different algorithms on BSDS500 and NASA dataset images. Significant values are in [bold]. Table 8Comparison of EPI values across different algorithms on BSDS500 and NASA dataset images.BSDS500NASAImage namenTAOEOPSOWOAProposedImage namenTAHAAOEOPSOWOAProposed12009330.71460.73910.41350.7255 0.7497 Kizimen30.56110.48590.55750.25630.5250 0.5613 50.76570.74930.35000.7423 0.8335 50.69490.56090.64080.29520.5340 0.6929 80.85960.78350.43530.7598 0.8858 80.76970.69930.59330.29130.5322 0.7727 15900230.47990.58720.34010.5961 0.6310 Yellowstone30.63040.46800.58380.29380.5868 0.6287 50.57130.65310.31650.5805 0.7460 50.73830.65520.65450.32360.6861 0.7401 80.77620.70270.38810.6404 0.8299 80.82610.76850.66020.38000.5862 0.8271 18902930.55970.57800.43090.5800 0.5960 Goldpanfire30.58540.43790.51410.44570.4822 0.5854 50.62960.62870.36900.5780 0.7843 50.68120.58950.57970.19670.6043 0.6816 80.63760.62310.35380.6611 0.8503 80.78680.69120.54770.21820.5729 0.7878 22807630.59200.68620.40310.6809 0.7069 Nyamuragira30.30120.25720.26800.22600.2577 0.3061 50.75040.76490.39260.7489 0.8191 50.43270.29270.30270.18870.2752 0.4303 80.75910.77190.42610.7203 0.8745 80.56090.44280.31730.19190.3117 0.5704 24700330.44920.48610.30020.4391 0.5276 Rockhouse30.41290.36770.36030.25140.3974 0.4288 50.51690.58090.31980.5189 0.6619 50.67120.50840.53780.25390.5579 0.6672 80.65100.58540.32080.5441 0.7809 80.74100.56040.48760.14350.3655 0.7410 32608530.38450.52360.39270.5175 0.5861 Yarrabin30.33800.28170.31330.23290.3435 0.3380 50.58870.58440.34530.5916 0.7029 50.55030.27600.31880.13830.3015 0.5526 80.67460.63280.08730.5536 0.7957 80.68190.47370.42500.14530.4233 0.6851 Significant values are in [bold]. Table 8Comparison of EPI values across different algorithms on BSDS500 and NASA dataset images.BSDS500NASAImage namenTAOEOPSOWOAProposedImage namenTAHAAOEOPSOWOAProposed12009330.71460.73910.41350.7255 0.7497 Kizimen30.56110.48590.55750.25630.5250 0.5613 50.76570.74930.35000.7423 0.8335 50.69490.56090.64080.29520.5340 0.6929 80.85960.78350.43530.7598 0.8858 80.76970.69930.59330.29130.5322 0.7727 15900230.47990.58720.34010.5961 0.6310 Yellowstone30.63040.46800.58380.29380.5868 0.6287 50.57130.65310.31650.5805 0.7460 50.73830.65520.65450.32360.6861 0.7401 80.77620.70270.38810.6404 0.8299 80.82610.76850.66020.38000.5862 0.8271 18902930.55970.57800.43090.5800 0.5960 Goldpanfire30.58540.43790.51410.44570.4822 0.5854 50.62960.62870.36900.5780 0.7843 50.68120.58950.57970.19670.6043 0.6816 80.63760.62310.35380.6611 0.8503 80.78680.69120.54770.21820.5729 0.7878 22807630.59200.68620.40310.6809 0.7069 Nyamuragira30.30120.25720.26800.22600.2577 0.3061 50.75040.76490.39260.7489 0.8191 50.43270.29270.30270.18870.2752 0.4303 80.75910.77190.42610.7203 0.8745 80.56090.44280.31730.19190.3117 0.5704 24700330.44920.48610.30020.4391 0.5276 Rockhouse30.41290.36770.36030.25140.3974 0.4288 50.51690.58090.31980.5189 0.6619 50.67120.50840.53780.25390.5579 0.6672 80.65100.58540.32080.5441 0.7809 80.74100.56040.48760.14350.3655 0.7410 32608530.38450.52360.39270.5175 0.5861 Yarrabin30.33800.28170.31330.23290.3435 0.3380 50.58870.58440.34530.5916 0.7029 50.55030.27600.31880.13830.3015 0.5526 80.67460.63280.08730.5536 0.7957 80.68190.47370.42500.14530.4233 0.6851 Significant values are in [bold]. Comparison of EPI values across different algorithms on BSDS500 and NASA dataset images. Significant values are in [bold]. Fig. 10Comparison of AO, EO, PSO, WOA, and Proposed method on BSDS500 228076 image across different threshold levels. Comparison of AO, EO, PSO, WOA, and Proposed method on BSDS500 228076 image across different threshold levels. Table 9Comparison of MIF values across different algorithms on BSDS500 and NASA dataset images.BSDS500NASAImage namenTAOEOPSOWOAProposedImage namenTAHAAOEOPSOWOAProposed12009331.89121.70550.95991.5816 1.7488 Kizimen31.74631.64691.75480.94021.6925 1.7646 52.08902.04201.07071.9683 2.1678 52.28562.16032.08130.91881.8907 2.2863 82.60512.09851.48721.9912 2.7443 82.77202.60662.10590.89961.9315 2.7760 15900231.91461.92151.49431.9171 1.9215 Yellowstone31.67991.70771.79111.20091.7054 1.7911 52.32972.27351.69242.2526 2.3341 52.15312.13492.12591.2571 2.1796 2.152582.63762.34041.73452.4022 2.8271 82.67712.42812.32501.20302.0639 2.6866 18902931.82871.83201.41791.8179 1.8320 Goldpanfire31.70401.6760 1.8162 1.66401.81451.704052.10662.09491.41012.0635 2.2006 52.24782.00061.99980.62711.9943 2.2512 82.39802.34301.45032.1240 2.6622 82.71312.51072.22920.80972.0769 2.7915 22807631.67701.82121.09391.8239 1.8239 Nyamuragira31.67921.74991.67581.23501.4861 1.8679 52.10112.11861.13692.1081 2.1528 52.37452.22491.93830.72951.7529 2.3745 82.45482.37321.17132.2461 2.6979 82.96142.82182.09840.92111.9502 2.9614 24700331.82571.75751.38661.5812 1.8038 Rockhouse31.69091.64401.70790.95251.6658 1.7195 52.32942.11071.59631.9908 2.2906 52.23172.02771.94540.96271.9469 2.2317 82.73672.20031.50732.1092 2.8566 82.77072.70121.96190.41571.6549 2.8038 32608531.82861.88721.41811.8844 1.9441 Yarrabin31.59291.7140 1.7702 1.13081.68301.592952.31422.30731.53702.2796 2.4912 52.06252.17952.05830.49661.8071 2.2103 82.77442.37820.71562.4622 3.0246 82.75222.54002.18560.72332.1688 2.7622 Significant values are in [bold]. Table 9Comparison of MIF values across different algorithms on BSDS500 and NASA dataset images.BSDS500NASAImage namenTAOEOPSOWOAProposedImage namenTAHAAOEOPSOWOAProposed12009331.89121.70550.95991.5816 1.7488 Kizimen31.74631.64691.75480.94021.6925 1.7646 52.08902.04201.07071.9683 2.1678 52.28562.16032.08130.91881.8907 2.2863 82.60512.09851.48721.9912 2.7443 82.77202.60662.10590.89961.9315 2.7760 15900231.91461.92151.49431.9171 1.9215 Yellowstone31.67991.70771.79111.20091.7054 1.7911 52.32972.27351.69242.2526 2.3341 52.15312.13492.12591.2571 2.1796 2.152582.63762.34041.73452.4022 2.8271 82.67712.42812.32501.20302.0639 2.6866 18902931.82871.83201.41791.8179 1.8320 Goldpanfire31.70401.6760 1.8162 1.66401.81451.704052.10662.09491.41012.0635 2.2006 52.24782.00061.99980.62711.9943 2.2512 82.39802.34301.45032.1240 2.6622 82.71312.51072.22920.80972.0769 2.7915 22807631.67701.82121.09391.8239 1.8239 Nyamuragira31.67921.74991.67581.23501.4861 1.8679 52.10112.11861.13692.1081 2.1528 52.37452.22491.93830.72951.7529 2.3745 82.45482.37321.17132.2461 2.6979 82.96142.82182.09840.92111.9502 2.9614 24700331.82571.75751.38661.5812 1.8038 Rockhouse31.69091.64401.70790.95251.6658 1.7195 52.32942.11071.59631.9908 2.2906 52.23172.02771.94540.96271.9469 2.2317 82.73672.20031.50732.1092 2.8566 82.77072.70121.96190.41571.6549 2.8038 32608531.82861.88721.41811.8844 1.9441 Yarrabin31.59291.7140 1.7702 1.13081.68301.592952.31422.30731.53702.2796 2.4912 52.06252.17952.05830.49661.8071 2.2103 82.77442.37820.71562.4622 3.0246 82.75222.54002.18560.72332.1688 2.7622 Significant values are in [bold]. Comparison of MIF values across different algorithms on BSDS500 and NASA dataset images. Significant values are in [bold]. Table 10Performance comparison of PSNR/SSIM/FSIM/QILV/CORR/EPI/MIF across different optimization techniques for BSDS dataset.AOEOPSOWOAProposedAO_0.8/0.8/1/0.8/0.9/0.8/11/1/1/1/1/1/10.9/0.9/1/1/1/0.9/10.0/0/0/0.2/0/0/0EO0.3/0.2/0/0.167/0.1/0.2/0_1/1/1/1/1/1/10.8/0.9/0.8/0.8/0.8/0.8/0.80.0/0/0/0/0/0/0PSO0.0/0/0/0/0/0/00/0/0/0/0/0/0_0.083/0/0/0/0/0/00/0/0/0/0/0/0WOA0.1/0.08/0/0/0/0.1/00.16/0.08/0.2/0.2/0.2/0.2/0.20.9/1/1/1/1/1/1_0/0/0/0/0/0/0Proposed1/1/1/0.8/1/1/11/1/1/1/1/1/11/1/1/1/1/1/11/1/1/1/1/1/1_ Table 10Performance comparison of PSNR/SSIM/FSIM/QILV/CORR/EPI/MIF across different optimization techniques for BSDS dataset.AOEOPSOWOAProposedAO_0.8/0.8/1/0.8/0.9/0.8/11/1/1/1/1/1/10.9/0.9/1/1/1/0.9/10.0/0/0/0.2/0/0/0EO0.3/0.2/0/0.167/0.1/0.2/0_1/1/1/1/1/1/10.8/0.9/0.8/0.8/0.8/0.8/0.80.0/0/0/0/0/0/0PSO0.0/0/0/0/0/0/00/0/0/0/0/0/0_0.083/0/0/0/0/0/00/0/0/0/0/0/0WOA0.1/0.08/0/0/0/0.1/00.16/0.08/0.2/0.2/0.2/0.2/0.20.9/1/1/1/1/1/1_0/0/0/0/0/0/0Proposed1/1/1/0.8/1/1/11/1/1/1/1/1/11/1/1/1/1/1/11/1/1/1/1/1/1_ Performance comparison of PSNR/SSIM/FSIM/QILV/CORR/EPI/MIF across different optimization techniques for BSDS dataset. Table 11Performance comparison of PSNR/SSIM/FSIM/QILV/CORR/EPI/MIF across different optimization techniques for NASA dataset.AOEOPSOWOAAHAProposedAO–0.5/0.8/0.7/0.8/0.6/0.4/0.90.1/1/1/1/1/1/10.6/0.8/0.8/0.8/0.7/0.4/0.80.1/0.3/0/0.4/0/0/0.30.1/0.3/0/0.4/0/0/0.3EO0.5/0.3/0.333/0.25/0.4/0.6/0.1–0.2/1/1/1/1/0.9/10.5/0.7/0.7/0.7/0.7/0.5/0.750.2/0.1/0.1/0.2/0/0/00.1/0.1/0.1/0.2/0/0/0PSO0.9/0/0/0/0/0/00.8/0/0/0/0/0.1/0–0.7/0/0/0.2/0/0/00.6/0/0/0/0/0/00.6/0/0/0/0/0/0WOA0.4/0.3/0.2/0.2/0.3/0.6/0.20.5/0.3/0.3/0.3/0.3/0.5/0.30.3/1/1/0.8/1/1/1–0.3/0.1/0.1/0.1/0/0/0.10.3/0.1/0.1/0.2/0/0/0.1AHA0.9/0.7/1/0.6/1/1/0.80.8/0.9/0.9/0.8/1/1/10.4/1/1/1/1/1/10.8/0.9/0.9/0.9/1/1/0.9–0.2/0.5/0.3/0.4/0.3/0.6/0.3Proposed0.9/0.7/1/0.6/1/1/0.80.9/0.9/0.9/0.8/1/1/10.4/1/1/1/1/1/10.8/0.9/0.9/0.8/1/1/0.90.8/0.5/0.7/0.6/0.7/0.4/0.7– Table 11Performance comparison of PSNR/SSIM/FSIM/QILV/CORR/EPI/MIF across different optimization techniques for NASA dataset.AOEOPSOWOAAHAProposedAO–0.5/0.8/0.7/0.8/0.6/0.4/0.90.1/1/1/1/1/1/10.6/0.8/0.8/0.8/0.7/0.4/0.80.1/0.3/0/0.4/0/0/0.30.1/0.3/0/0.4/0/0/0.3EO0.5/0.3/0.333/0.25/0.4/0.6/0.1–0.2/1/1/1/1/0.9/10.5/0.7/0.7/0.7/0.7/0.5/0.750.2/0.1/0.1/0.2/0/0/00.1/0.1/0.1/0.2/0/0/0PSO0.9/0/0/0/0/0/00.8/0/0/0/0/0.1/0–0.7/0/0/0.2/0/0/00.6/0/0/0/0/0/00.6/0/0/0/0/0/0WOA0.4/0.3/0.2/0.2/0.3/0.6/0.20.5/0.3/0.3/0.3/0.3/0.5/0.30.3/1/1/0.8/1/1/1–0.3/0.1/0.1/0.1/0/0/0.10.3/0.1/0.1/0.2/0/0/0.1AHA0.9/0.7/1/0.6/1/1/0.80.8/0.9/0.9/0.8/1/1/10.4/1/1/1/1/1/10.8/0.9/0.9/0.9/1/1/0.9–0.2/0.5/0.3/0.4/0.3/0.6/0.3Proposed0.9/0.7/1/0.6/1/1/0.80.9/0.9/0.9/0.8/1/1/10.4/1/1/1/1/1/10.8/0.9/0.9/0.8/1/1/0.90.8/0.5/0.7/0.6/0.7/0.4/0.7– Performance comparison of PSNR/SSIM/FSIM/QILV/CORR/EPI/MIF across different optimization techniques for NASA dataset. 0.5/0.8/0.7/0.8/ 0.6/0.4/0.9 0.6/0.8/0.8/0.8/0.7 /0.4/0.8 0.5/0.3/0.333/0.25 /0.4/0.6/0.1 0.5/0.7/0.7/0.7/0.7 /0.5/0.75 0.4/0.3/0.2/0.2/0.3 /0.6/0.2 0.5/0.3/0.3/0.3/ 0.3/0.5/0.3 0.2/0.5/0.3/0.4/0.3 /0.6/0.3 Correlation coefficient The correlation coefficient was used to examine how closely the intensity pattern of the segmented image follows that of the original image. In threshold-based segmentation, a higher correlation value suggests that the overall intensity distribution and image structure are preserved with less distortion. Table 7 reports the correlation coefficient values for the BSDS500 and NASA images. The proposed method shows consistent correlation values across different threshold levels, indicating that the segmented outputs remain closely related to the original images. This behaviour is useful for both natural and satellite images, where maintaining the overall intensity pattern is important for reliable segmentation. Edge preservation index Edge preservation is an important requirement in multilevel segmentation, as poorly selected thresholds may blur or break object boundaries. EPI was therefore used to compare the edge information retained in the segmented image with that of the original image. A higher EPI value indicates better preservation of important boundary and structural details. The EPI values in Table 8 show that the proposed method preserves edge information more reliably across the tested images. The improvement is more noticeable at higher threshold levels, where some baseline algorithms lose boundary clarity. This result is important for both natural images and satellite images, since clear edges help maintain object separation and spatial details after segmentation.Table 12Performance comparison of multilevel RGB threshold values on BSDS500 dataset using various optimization techniques.ImagenamenTAOEOPSOWOAProposed methodRGBRGBRGBRGBRGB120093369 125 18958 118 19433 70 12063 103 18133 93 16332 77 14163 103 18133 93 16231 76 14120 77 15933 92 16235 77 14166 119 17964 116 17446 92 146546 46 86 96 17853 107 116 159 20023 32 81 120 1601 22 65 104 1802 27 58 97 16510 39 68 114 14923 62 96 142 19629 63 97 140 19830 71 110 143 18111 41 60 110 1951 20 63 96 1573 32 63 86 14637 72 110 152 19736 71 106 146 19230 60 93 132 175828 31 60 81 85 119 158 20926 55 59 104 124 159 187 1879 30 71 109 132 137 138 2461 1 4 24 63 102 172 2551 17 31 59 94 145 218 2551 2 10 35 71 114 150 25538 55 71 103 113 136 159 18757 90 120 127 153 188 220 25510 40 81 102 115 143 171 1961 1 1 2 8 65 108 1871 1 1 1 1 38 94 1611 1 2 3 22 68 111 15126 51 75 99 126 150 182 21228 52 74 99 126 152 181 21221 41 62 85 107 135 166 205159002328 72 12422 100 15439 99 24927 85 16427 80 15622 57 11527 85 16427 79 15422 57 11529 82 16427 78 15421 59 11648 102 16750 100 16130 73 134514 40 86 126 22324 70 108 144 15726 57 101 159 1627 26 67 113 1787 27 51 92 1505 20 43 75 12326 65 106 160 21122 55 92 138 18719 43 75 111 15215 22 76 118 1802 5 32 85 1582 20 45 82 12730 62 99 140 19134 66 100 138 18721 48 81 125 183815 21 52 77 106 136 162 21316 35 53 89 97 159 187 21611 14 34 51 93 125 229 2401 1 5 20 30 63 120 1781 1 1 10 26 72 113 1821 1 1 5 18 40 71 12224 60 79 105 122 129 184 1846 26 60 100 151 164 204 25519 43 76 106 142 177 241 2551 1 1 25 39 54 99 1761 1 1 1 2 33 74 15922 58 115 240 247 253 255 25521 42 64 87 114 143 179 21122 40 63 87 108 136 171 20915 28 49 71 95 125 155 196189029317 88 18331 108 19029 79 18929 96 19127 90 18528 80 14932 99 19227 90 18528 80 14932 99 19134 97 18627 80 15251 112 18148 105 17437 87 151518 52 136 208 22918 66 123 169 18127 61 126 146 1936 25 71 125 19922 47 81 136 2084 23 55 103 16221 57 103 160 21821 46 90 147 20920 42 74 117 16928 51 114 156 22027 84 146 194 20825 56 71 116 17932 67 109 154 20532 66 107 151 20025 54 91 134 183817 39 39 46 57 115 163 23224 53 64 109 144 164 205 25028 42 57 59 70 105 132 1661 1 4 29 61 110 187 2091 1 4 23 41 79 165 2075 22 42 68 109 164 228 24225 42 61 101 126 128 163 19916 25 39 66 85 106 142 18020 31 50 96 108 155 186 2146 20 49 100 171 232 232 2322 22 49 67 110 153 163 2151 1 5 21 57 82 109 16720 38 65 94 125 157 190 22522 42 65 92 122 154 186 21816 35 57 82 114 141 176 209228076329 76 12824 54 11237 75 15018 56 11521 61 12129 80 17025 74 21221 61 12129 80 17018 55 1055 32 10227 79 17035 82 15139 86 15039 86 156525 68 94 126 21729 83 143 157 2145 21 62 125 1926 20 44 75 1411 24 63 120 2224 25 57 105 18415 42 75 119 21221 52 91 132 16316 36 65 109 18920 53 73 119 18724 49 87 129 23011 31 50 106 19022 48 81 121 17727 56 91 134 19024 50 81 123 17988 20 45 62 109 133 180 23019 57 66 128 161 197 239 2465 25 45 56 76 93 101 1461 4 17 32 57 94 186 2551 1 8 26 58 107 150 2221 4 20 47 76 104 185 2555 25 48 70 95 132 212 23520 51 79 117 151 169 219 22420 42 73 96 127 170 212 2351 1 1 3 27 48 91 1331 1 1 2 3 27 62 1196 16 60 102 185 207 255 25514 28 43 61 81 112 151 19118 35 51 71 93 120 156 20315 29 49 69 92 125 161 197247003362 109 18841 101 15930 55 10922 69 13636 94 15444 76 12622 69 13736 93 15344 76 12626 69 13538 91 15343 75 12554 99 16356 105 16350 91 150538 67 77 132 16859 73 96 146 21844 69 80 136 2524 21 57 90 1481 22 76 109 15814 37 56 88 13228 63 86 115 16019 56 87 122 17431 57 89 136 2516 26 58 107 1706 50 94 131 19810 26 53 96 14333 63 95 138 19036 72 103 140 18931 56 87 126 179826 55 80 107 144 166 169 23537 59 94 120 143 147 187 25330 65 71 91 129 185 215 2181 1 5 12 24 62 99 1541 4 31 61 95 149 255 2551 1 2 22 48 63 98 13733 62 80 102 132 136 179 22410 17 25 71 106 138 185 24817 43 64 90 123 166 171 21222 50 117 187 187 187 187 18712 33 87 133 238 255 255 2557 18 31 51 93 107 107 12222 42 61 80 101 130 166 20723 49 70 92 114 142 176 21023 43 62 80 102 128 160 201326085370 103 16950 88 17837 64 14639 103 17629 88 16727 71 14939 102 17529 88 16627 71 14943 97 16831 88 16227 71 14957 110 17150 98 16241 83 151537 79 120 170 20427 38 98 135 17813 37 67 188 20315 30 75 124 1742 26 71 111 17914 38 63 90 16226 67 106 157 20323 60 91 125 18722 53 83 112 17324 43 85 139 18632 52 93 121 19519 49 80 120 15738 73 107 145 19431 64 95 131 18525 52 82 121 178830 49 62 104 108 142 185 23715 33 62 100 143 146 197 2347 39 70 92 108 135 152 1721 17 20 22 41 83 125 1741 3 16 33 64 102 129 1771 1 4 17 46 71 100 16420 32 45 90 137 156 176 19923 53 72 90 109 128 151 18716 43 74 100 131 169 174 1851 18 52 96 160 205 205 2051 1 10 28 53 93 138 2031 10 18 58 91 95 125 17823 46 69 90 118 142 173 21021 43 63 81 104 129 158 20419 33 50 67 90 117 156 201Table 13Performance comparison of multilevel RGB threshold values on NASA dataset using various optimization techniques.ImagenamenTAHAAOEOPSOWOAProposed methodRGBRGBRGBRGBRGBRGBKizimen377 142 19455 100 15040 83 13472 141 21056 102 16541 62 9571 153 20353 99 14827 73 12269 141 19654 100 15033 76 12469 132 19951 87 14528 69 12277 142 19455 100 15040 83 134552 97 141 177 20939 70 103 139 17426 53 85 120 16217 78 144 183 19664 96 129 172 21312 54 76 94 1302 74 141 196 25112 55 88 117 16116 42 74 106 14052 73 95 109 17621 66 92 126 15822 59 99 138 23169 86 107 173 1912 28 82 103 1502 2 35 77 12852 99 143 179 21039 70 103 140 17526 52 84 118 162826 58 91 122 151 176 199 22225 47 68 91 115 140 164 19417 35 54 76 100 124 150 18943 53 81 119 133 173 209 24021 49 63 91 115 144 157 24518 50 54 97 120 134 148 1761 1 1 2 11 69 141 1841 7 9 47 94 115 136 1713 8 16 17 39 68 96 13354 62 74 83 91 92 109 18420 21 22 51 75 121 159 24628 47 67 69 76 126 225 2371 2 9 13 21 73 115 19727 39 51 95 147 184 255 2552 3 24 56 83 134 134 13423 53 83 113 149 175 198 21928 48 69 92 119 141 163 19420 39 59 81 103 126 153 200Yellowstone328 78 14364 117 15547 78 14015 38 14846 113 14115 48 7411 55 13652 114 14828 49 7714 61 14171 117 14944 68 9419 68 13862 121 1445 47 7628 78 14365 117 15547 78 140516 43 78 121 17145 89 121 149 18631 51 72 98 15415 65 120 151 18159 117 153 172 20434 39 57 82 1796 30 56 102 1662 35 96 127 15511 41 63 90 2506 18 51 117 15814 35 91 120 15113 44 66 92 2026 22 57 131 18929 78 106 143 17112 22 45 73 9716 44 81 124 17545 89 121 150 18731 50 72 99 153810 25 44 67 93 122 155 19227 59 88 110 128 147 167 20121 37 51 66 82 101 134 1888 39 77 83 119 136 195 2123 50 55 81 119 147 170 17627 42 59 90 104 143 149 2051 2 5 18 45 109 165 2551 1 6 41 58 111 138 1591 7 17 39 65 89 176 23210 42 67 81 84 122 149 17656 77 110 116 125 139 144 18217 46 71 100 142 189 195 2175 17 17 63 133 162 255 2555 16 32 117 146 240 255 2551 1 2 5 17 41 63 909 22 40 59 89 118 154 18927 59 89 105 123 142 160 19624 38 53 68 84 105 138 195Goldpanfire353 103 15874 127 17257 110 14955 89 166101 144 20990 109 16642 95 15785 136 18276 106 13246 99 16087 132 17696 119 14565 115 16029 108 165106 119 13653 103 15874 127 17257 110 149534 67 101 138 18046 89 124 156 18947 91 115 139 18233 68 123 166 22817 80 123 150 16929 45 55 102 1347 23 70 102 1612 35 95 143 1831 17 94 118 1429 27 60 115 17560 94 119 142 18147 96 115 136 18222 76 132 182 25550 91 128 177 17712 38 95 118 14134 67 102 138 18147 89 126 157 18945 91 115 139 182823 45 67 90 114 140 167 19929 61 89 113 135 157 179 20428 61 86 103 119 135 155 19412 28 60 88 113 154 221 23439 77 96 121 131 161 168 1812 29 69 83 106 130 154 1876 8 19 32 75 97 118 1701 1 2 34 104 120 147 1851 2 6 14 36 95 118 1446 35 68 91 141 144 166 20454 66 81 89 97 133 168 25534 58 107 139 201 204 212 22636 61 108 123 170 255 255 2551 1 7 17 83 115 138 1981 23 30 35 38 90 115 14623 47 66 87 111 136 168 20229 61 89 117 136 162 181 20527 58 87 102 121 137 158 196Nyamuragira365 116 16658 121 17859 108 16628 115 14519 138 18744 91 1569 88 15010 122 18010 78 14614 88 14823 123 17912 78 1461 19 11934 118 18121 78 14865 116 16658 121 17859 108 166535 76 112 147 18534 77 116 151 19430 66 102 141 18928 58 105 136 15010 115 136 167 21915 27 58 102 18016 66 101 156 2553 29 115 173 25512 16 64 107 16120 49 79 119 18433 57 102 138 19113 61 101 133 1811 2 63 83 1543 15 60 122 1891 1 15 81 14335 76 112 147 18534 77 116 151 19430 66 102 141 189826 55 80 104 128 152 177 20722 51 81 107 131 154 182 21623 48 71 94 117 143 173 2109 54 77 96 120 145 186 2352 31 92 121 147 165 196 2086 38 55 104 124 166 192 2351 2 11 75 116 162 254 2551 1 1 1 3 114 135 1821 1 5 17 63 101 149 25534 43 82 92 130 188 189 23129 36 42 56 59 72 115 16328 41 45 51 78 137 173 2553 8 34 82 136 163 250 2501 1 1 1 2 42 115 1823 61 110 167 255 255 255 25526 55 80 104 128 152 177 20722 51 81 107 131 154 182 21623 48 71 94 117 143 173 210Rockhouse366 120 17366 113 16165 113 162104 146 16799 101 15481 132 19685 126 17967 104 15962 96 15890 137 18375 121 16669 111 1613 115 17644 98 16279 141 17066 120 17366 113 16165 113 162539 80 117 154 18841 76 110 143 17841 77 111 145 18068 109 145 177 24712 66 112 169 24741 88 127 163 19025 79 113 152 19313 25 72 123 1673 51 85 115 15526 78 114 154 19366 92 120 164 23421 67 107 147 1792 50 99 135 1771 18 74 118 16432 61 106 142 18139 80 117 154 18841 76 110 143 17841 77 111 145 180824 52 77 101 128 154 178 20226 52 75 99 123 147 171 19926 53 77 100 124 148 173 20021 32 73 108 120 153 178 20155 63 75 85 117 149 190 23066 92 111 148 169 183 190 1971 1 1 2 27 99 147 1861 1 6 11 27 70 116 1661 1 4 13 64 99 154 18722 41 80 85 108 130 181 24736 39 49 67 98 138 181 24169 77 116 119 125 156 200 2457 7 20 30 59 136 136 17350 83 160 160 160 160 160 1603 82 136 153 154 154 154 15424 52 77 101 128 154 178 20226 52 75 99 123 147 171 19926 53 77 100 124 148 173 200Yarrabin338 97 16756 112 16851 105 15812 93 16567 123 18081 112 17817 88 17053 110 16893 122 15621 90 17164 118 17089 115 14835 72 15676 132 17423 100 13538 97 16756 112 16851 105 158522 57 99 146 19535 74 113 150 18741 83 108 136 18127 55 135 167 20947 92 97 134 18175 96 119 172 2321 11 40 113 18520 42 95 140 1801 15 87 112 14514 27 44 110 17633 84 127 160 19541 75 94 118 1521 18 58 136 1871 2 42 108 1661 1 13 100 13722 57 99 146 19535 74 113 150 18741 83 108 136 181812 30 53 80 110 142 176 21221 46 71 98 125 150 175 20322 52 78 95 112 131 157 1988 27 88 141 187 206 226 2395 26 63 84 120 149 187 25411 91 111 130 171 205 212 2231 1 3 17 45 103 157 2091 2 3 11 43 95 132 1725 12 28 87 113 147 248 2549 30 43 46 55 76 115 16758 81 98 110 112 130 149 19423 26 31 49 83 107 128 1591 1 11 32 89 149 203 25524 34 75 97 111 129 170 25596 129 183 238 255 255 255 25512 30 53 80 110 142 176 21221 46 71 98 125 150 175 20322 52 78 95 112 131 157 198 Table 12Performance comparison of multilevel RGB threshold values on BSDS500 dataset using various optimization techniques.ImagenamenTAOEOPSOWOAProposed methodRGBRGBRGBRGBRGB120093369 125 18958 118 19433 70 12063 103 18133 93 16332 77 14163 103 18133 93 16231 76 14120 77 15933 92 16235 77 14166 119 17964 116 17446 92 146546 46 86 96 17853 107 116 159 20023 32 81 120 1601 22 65 104 1802 27 58 97 16510 39 68 114 14923 62 96 142 19629 63 97 140 19830 71 110 143 18111 41 60 110 1951 20 63 96 1573 32 63 86 14637 72 110 152 19736 71 106 146 19230 60 93 132 175828 31 60 81 85 119 158 20926 55 59 104 124 159 187 1879 30 71 109 132 137 138 2461 1 4 24 63 102 172 2551 17 31 59 94 145 218 2551 2 10 35 71 114 150 25538 55 71 103 113 136 159 18757 90 120 127 153 188 220 25510 40 81 102 115 143 171 1961 1 1 2 8 65 108 1871 1 1 1 1 38 94 1611 1 2 3 22 68 111 15126 51 75 99 126 150 182 21228 52 74 99 126 152 181 21221 41 62 85 107 135 166 205159002328 72 12422 100 15439 99 24927 85 16427 80 15622 57 11527 85 16427 79 15422 57 11529 82 16427 78 15421 59 11648 102 16750 100 16130 73 134514 40 86 126 22324 70 108 144 15726 57 101 159 1627 26 67 113 1787 27 51 92 1505 20 43 75 12326 65 106 160 21122 55 92 138 18719 43 75 111 15215 22 76 118 1802 5 32 85 1582 20 45 82 12730 62 99 140 19134 66 100 138 18721 48 81 125 183815 21 52 77 106 136 162 21316 35 53 89 97 159 187 21611 14 34 51 93 125 229 2401 1 5 20 30 63 120 1781 1 1 10 26 72 113 1821 1 1 5 18 40 71 12224 60 79 105 122 129 184 1846 26 60 100 151 164 204 25519 43 76 106 142 177 241 2551 1 1 25 39 54 99 1761 1 1 1 2 33 74 15922 58 115 240 247 253 255 25521 42 64 87 114 143 179 21122 40 63 87 108 136 171 20915 28 49 71 95 125 155 196189029317 88 18331 108 19029 79 18929 96 19127 90 18528 80 14932 99 19227 90 18528 80 14932 99 19134 97 18627 80 15251 112 18148 105 17437 87 151518 52 136 208 22918 66 123 169 18127 61 126 146 1936 25 71 125 19922 47 81 136 2084 23 55 103 16221 57 103 160 21821 46 90 147 20920 42 74 117 16928 51 114 156 22027 84 146 194 20825 56 71 116 17932 67 109 154 20532 66 107 151 20025 54 91 134 183817 39 39 46 57 115 163 23224 53 64 109 144 164 205 25028 42 57 59 70 105 132 1661 1 4 29 61 110 187 2091 1 4 23 41 79 165 2075 22 42 68 109 164 228 24225 42 61 101 126 128 163 19916 25 39 66 85 106 142 18020 31 50 96 108 155 186 2146 20 49 100 171 232 232 2322 22 49 67 110 153 163 2151 1 5 21 57 82 109 16720 38 65 94 125 157 190 22522 42 65 92 122 154 186 21816 35 57 82 114 141 176 209228076329 76 12824 54 11237 75 15018 56 11521 61 12129 80 17025 74 21221 61 12129 80 17018 55 1055 32 10227 79 17035 82 15139 86 15039 86 156525 68 94 126 21729 83 143 157 2145 21 62 125 1926 20 44 75 1411 24 63 120 2224 25 57 105 18415 42 75 119 21221 52 91 132 16316 36 65 109 18920 53 73 119 18724 49 87 129 23011 31 50 106 19022 48 81 121 17727 56 91 134 19024 50 81 123 17988 20 45 62 109 133 180 23019 57 66 128 161 197 239 2465 25 45 56 76 93 101 1461 4 17 32 57 94 186 2551 1 8 26 58 107 150 2221 4 20 47 76 104 185 2555 25 48 70 95 132 212 23520 51 79 117 151 169 219 22420 42 73 96 127 170 212 2351 1 1 3 27 48 91 1331 1 1 2 3 27 62 1196 16 60 102 185 207 255 25514 28 43 61 81 112 151 19118 35 51 71 93 120 156 20315 29 49 69 92 125 161 197247003362 109 18841 101 15930 55 10922 69 13636 94 15444 76 12622 69 13736 93 15344 76 12626 69 13538 91 15343 75 12554 99 16356 105 16350 91 150538 67 77 132 16859 73 96 146 21844 69 80 136 2524 21 57 90 1481 22 76 109 15814 37 56 88 13228 63 86 115 16019 56 87 122 17431 57 89 136 2516 26 58 107 1706 50 94 131 19810 26 53 96 14333 63 95 138 19036 72 103 140 18931 56 87 126 179826 55 80 107 144 166 169 23537 59 94 120 143 147 187 25330 65 71 91 129 185 215 2181 1 5 12 24 62 99 1541 4 31 61 95 149 255 2551 1 2 22 48 63 98 13733 62 80 102 132 136 179 22410 17 25 71 106 138 185 24817 43 64 90 123 166 171 21222 50 117 187 187 187 187 18712 33 87 133 238 255 255 2557 18 31 51 93 107 107 12222 42 61 80 101 130 166 20723 49 70 92 114 142 176 21023 43 62 80 102 128 160 201326085370 103 16950 88 17837 64 14639 103 17629 88 16727 71 14939 102 17529 88 16627 71 14943 97 16831 88 16227 71 14957 110 17150 98 16241 83 151537 79 120 170 20427 38 98 135 17813 37 67 188 20315 30 75 124 1742 26 71 111 17914 38 63 90 16226 67 106 157 20323 60 91 125 18722 53 83 112 17324 43 85 139 18632 52 93 121 19519 49 80 120 15738 73 107 145 19431 64 95 131 18525 52 82 121 178830 49 62 104 108 142 185 23715 33 62 100 143 146 197 2347 39 70 92 108 135 152 1721 17 20 22 41 83 125 1741 3 16 33 64 102 129 1771 1 4 17 46 71 100 16420 32 45 90 137 156 176 19923 53 72 90 109 128 151 18716 43 74 100 131 169 174 1851 18 52 96 160 205 205 2051 1 10 28 53 93 138 2031 10 18 58 91 95 125 17823 46 69 90 118 142 173 21021 43 63 81 104 129 158 20419 33 50 67 90 117 156 201 Performance comparison of multilevel RGB threshold values on BSDS500 dataset using various optimization techniques. Table 13Performance comparison of multilevel RGB threshold values on NASA dataset using various optimization techniques.ImagenamenTAHAAOEOPSOWOAProposed methodRGBRGBRGBRGBRGBRGBKizimen377 142 19455 100 15040 83 13472 141 21056 102 16541 62 9571 153 20353 99 14827 73 12269 141 19654 100 15033 76 12469 132 19951 87 14528 69 12277 142 19455 100 15040 83 134552 97 141 177 20939 70 103 139 17426 53 85 120 16217 78 144 183 19664 96 129 172 21312 54 76 94 1302 74 141 196 25112 55 88 117 16116 42 74 106 14052 73 95 109 17621 66 92 126 15822 59 99 138 23169 86 107 173 1912 28 82 103 1502 2 35 77 12852 99 143 179 21039 70 103 140 17526 52 84 118 162826 58 91 122 151 176 199 22225 47 68 91 115 140 164 19417 35 54 76 100 124 150 18943 53 81 119 133 173 209 24021 49 63 91 115 144 157 24518 50 54 97 120 134 148 1761 1 1 2 11 69 141 1841 7 9 47 94 115 136 1713 8 16 17 39 68 96 13354 62 74 83 91 92 109 18420 21 22 51 75 121 159 24628 47 67 69 76 126 225 2371 2 9 13 21 73 115 19727 39 51 95 147 184 255 2552 3 24 56 83 134 134 13423 53 83 113 149 175 198 21928 48 69 92 119 141 163 19420 39 59 81 103 126 153 200Yellowstone328 78 14364 117 15547 78 14015 38 14846 113 14115 48 7411 55 13652 114 14828 49 7714 61 14171 117 14944 68 9419 68 13862 121 1445 47 7628 78 14365 117 15547 78 140516 43 78 121 17145 89 121 149 18631 51 72 98 15415 65 120 151 18159 117 153 172 20434 39 57 82 1796 30 56 102 1662 35 96 127 15511 41 63 90 2506 18 51 117 15814 35 91 120 15113 44 66 92 2026 22 57 131 18929 78 106 143 17112 22 45 73 9716 44 81 124 17545 89 121 150 18731 50 72 99 153810 25 44 67 93 122 155 19227 59 88 110 128 147 167 20121 37 51 66 82 101 134 1888 39 77 83 119 136 195 2123 50 55 81 119 147 170 17627 42 59 90 104 143 149 2051 2 5 18 45 109 165 2551 1 6 41 58 111 138 1591 7 17 39 65 89 176 23210 42 67 81 84 122 149 17656 77 110 116 125 139 144 18217 46 71 100 142 189 195 2175 17 17 63 133 162 255 2555 16 32 117 146 240 255 2551 1 2 5 17 41 63 909 22 40 59 89 118 154 18927 59 89 105 123 142 160 19624 38 53 68 84 105 138 195Goldpanfire353 103 15874 127 17257 110 14955 89 166101 144 20990 109 16642 95 15785 136 18276 106 13246 99 16087 132 17696 119 14565 115 16029 108 165106 119 13653 103 15874 127 17257 110 149534 67 101 138 18046 89 124 156 18947 91 115 139 18233 68 123 166 22817 80 123 150 16929 45 55 102 1347 23 70 102 1612 35 95 143 1831 17 94 118 1429 27 60 115 17560 94 119 142 18147 96 115 136 18222 76 132 182 25550 91 128 177 17712 38 95 118 14134 67 102 138 18147 89 126 157 18945 91 115 139 182823 45 67 90 114 140 167 19929 61 89 113 135 157 179 20428 61 86 103 119 135 155 19412 28 60 88 113 154 221 23439 77 96 121 131 161 168 1812 29 69 83 106 130 154 1876 8 19 32 75 97 118 1701 1 2 34 104 120 147 1851 2 6 14 36 95 118 1446 35 68 91 141 144 166 20454 66 81 89 97 133 168 25534 58 107 139 201 204 212 22636 61 108 123 170 255 255 2551 1 7 17 83 115 138 1981 23 30 35 38 90 115 14623 47 66 87 111 136 168 20229 61 89 117 136 162 181 20527 58 87 102 121 137 158 196Nyamuragira365 116 16658 121 17859 108 16628 115 14519 138 18744 91 1569 88 15010 122 18010 78 14614 88 14823 123 17912 78 1461 19 11934 118 18121 78 14865 116 16658 121 17859 108 166535 76 112 147 18534 77 116 151 19430 66 102 141 18928 58 105 136 15010 115 136 167 21915 27 58 102 18016 66 101 156 2553 29 115 173 25512 16 64 107 16120 49 79 119 18433 57 102 138 19113 61 101 133 1811 2 63 83 1543 15 60 122 1891 1 15 81 14335 76 112 147 18534 77 116 151 19430 66 102 141 189826 55 80 104 128 152 177 20722 51 81 107 131 154 182 21623 48 71 94 117 143 173 2109 54 77 96 120 145 186 2352 31 92 121 147 165 196 2086 38 55 104 124 166 192 2351 2 11 75 116 162 254 2551 1 1 1 3 114 135 1821 1 5 17 63 101 149 25534 43 82 92 130 188 189 23129 36 42 56 59 72 115 16328 41 45 51 78 137 173 2553 8 34 82 136 163 250 2501 1 1 1 2 42 115 1823 61 110 167 255 255 255 25526 55 80 104 128 152 177 20722 51 81 107 131 154 182 21623 48 71 94 117 143 173 210Rockhouse366 120 17366 113 16165 113 162104 146 16799 101 15481 132 19685 126 17967 104 15962 96 15890 137 18375 121 16669 111 1613 115 17644 98 16279 141 17066 120 17366 113 16165 113 162539 80 117 154 18841 76 110 143 17841 77 111 145 18068 109 145 177 24712 66 112 169 24741 88 127 163 19025 79 113 152 19313 25 72 123 1673 51 85 115 15526 78 114 154 19366 92 120 164 23421 67 107 147 1792 50 99 135 1771 18 74 118 16432 61 106 142 18139 80 117 154 18841 76 110 143 17841 77 111 145 180824 52 77 101 128 154 178 20226 52 75 99 123 147 171 19926 53 77 100 124 148 173 20021 32 73 108 120 153 178 20155 63 75 85 117 149 190 23066 92 111 148 169 183 190 1971 1 1 2 27 99 147 1861 1 6 11 27 70 116 1661 1 4 13 64 99 154 18722 41 80 85 108 130 181 24736 39 49 67 98 138 181 24169 77 116 119 125 156 200 2457 7 20 30 59 136 136 17350 83 160 160 160 160 160 1603 82 136 153 154 154 154 15424 52 77 101 128 154 178 20226 52 75 99 123 147 171 19926 53 77 100 124 148 173 200Yarrabin338 97 16756 112 16851 105 15812 93 16567 123 18081 112 17817 88 17053 110 16893 122 15621 90 17164 118 17089 115 14835 72 15676 132 17423 100 13538 97 16756 112 16851 105 158522 57 99 146 19535 74 113 150 18741 83 108 136 18127 55 135 167 20947 92 97 134 18175 96 119 172 2321 11 40 113 18520 42 95 140 1801 15 87 112 14514 27 44 110 17633 84 127 160 19541 75 94 118 1521 18 58 136 1871 2 42 108 1661 1 13 100 13722 57 99 146 19535 74 113 150 18741 83 108 136 181812 30 53 80 110 142 176 21221 46 71 98 125 150 175 20322 52 78 95 112 131 157 1988 27 88 141 187 206 226 2395 26 63 84 120 149 187 25411 91 111 130 171 205 212 2231 1 3 17 45 103 157 2091 2 3 11 43 95 132 1725 12 28 87 113 147 248 2549 30 43 46 55 76 115 16758 81 98 110 112 130 149 19423 26 31 49 83 107 128 1591 1 11 32 89 149 203 25524 34 75 97 111 129 170 25596 129 183 238 255 255 255 25512 30 53 80 110 142 176 21221 46 71 98 125 150 175 20322 52 78 95 112 131 157 198 Performance comparison of multilevel RGB threshold values on NASA dataset using various optimization techniques. Fig. 11Comparison of AO, EO, PSO, WOA, and Proposed method on NASA image Goldpanfire across different threshold levels. Comparison of AO, EO, PSO, WOA, and Proposed method on NASA image Goldpanfire across different threshold levels. Mutual information factor MIF was used to analyze the extent of information conservation in the segmented image compared to the original one. For multilevel thresholding, an increase in MIF signifies conservation of information concerning intensity and loss of less information through segmentation. From Table 9, it is evident that the MIF values confirm the validity of the proposed technique. Compared to the baseline techniques, the proposed algorithm better conserves the image information in most cases across the test datasets of BSDS500 and NASA images. This is particularly significant when dealing with complicated images since too much thresholding tends to destroy minute details in the image. Comparison algorithms channel wise threshold distribution Threshold values obtained for the RGB color components of the BSDS500 and NASA datasets are illustrated in Tables 12 and 13. Thresholds display variations for each color component owing to differing properties such as light intensities, textures, and contents of the images. In comparison with conventional techniques, the threshold sets obtained using the proposed technique are more consistent especially for high threshold values, since the search space is harder in such cases. Pairwise statistical comparison of competing algorithms A pairwise comparison was also carried out using PSNR, SSIM, FSIM, QILV, CORR, EPI, and MIF to give a broader view of segmentation quality. Tables 10 and 11 summarize the results for the BSDS500 and NASA datasets, respectively. Overall, the proposed method shows more consistent performance than the baseline algorithms across most of the evaluated metrics. The results suggest that the method is able to preserve reconstruction quality, structural details, edge information, and useful image content across both natural and satellite images. Qualitative evaluation A qualitative visual comparison was also conducted to further evaluate the segmentation performance besides the quantitative evaluation. The representative results of various images are provided to emphasize clarity of boundaries, consistency of regions and maintenance of fine details. Figure 6 indicates that at an eight level threshold, image 247003 in the BSDS500 dataset, the proposed method exhibits better results in terms of boundary preservation and region refinement than AO, EO, PSO, and WOA. Figure 7 shows how the selected threshold vectors partition the histogram and energy curve of the BSDS500 image into nT+1 regions, supporting the visual segmentation result. The Yellowstone satellite image in Fig. 8 represents the image at third level threshold, indicating that the proposed method will consistently retain the structural information and minimize the over segmentation. Figure 9 presents the same threshold partitioning for the NASA Goldpanfire image, showing how the energy curve helps maintain stable region separation at the selected threshold level. The comparative outcomes in various threshold levels of BSDS500 and NASA images, in Figs. 10 and 11 respectively, confirm that the proposed approach is always able to draw sharper boundaries and higher homogeneity in comparison to the existing methods. These visual observations also supplement the quantitative findings, supporting the high efficiency and versatility of the proposed segmentation method in natural and satellite images. Visual comparison In order to evaluate the ability of the proposed technique in retaining edges, magnified images from difficult areas containing complex textures and weak boundaries are shown in the red boxed area in Fig. 12. In these difficult regions, it has been observed that the conventional metaheuristics algorithms, WOA and PSO, cause heavy edge distortion or over-segmentation. The new model, on the other hand, maintains sharp edges throughout. The superiority of the method is highlighted even more when used on architectural images, where the retention of the edge between the building and background is much better than AO and EO approaches.Fig. 12Visual comparison of edge preservation with zoomed rectangular insets. The central region of each image is enlarged inside the rectangular inset to highlight differences in boundary continuity. Visual comparison of edge preservation with zoomed rectangular insets. The central region of each image is enlarged inside the rectangular inset to highlight differences in boundary continuity. Ablation study of initialization strategies In order to clarify the significance of each mathematical sequence used during the initialization process, an ablation study is conducted. All Latin Hypercube, Sobol, Halton, and Sierpinski algorithms have been tested separately for one complex natural image from the BSDS500 dataset (Image 120093) with the usage of an 8-level thresholding. In order to make results comparable, the same number of iterations and population sizes were used for all methods. The values of convergence time and the resulting fitness are shown in Table 14. It can be concluded that the usage of only one initialization sequence limits the algorithm significantly. For instance, the usage of the Latin Hypercube strategy leads to the shortest convergence time (2.19 seconds), although its fitness value equals to 0.3615 and means the tendency to converge into a local minimum. Moreover, even though both Sobol and Halton sequences require a greater amount of time to execute, they still do not guarantee the highest accuracy of the algorithm. By combining four mathematical sequences together, the suggested algorithm ensures both a reasonable execution time of 3.86 seconds and the highest fitness value of 0.3495.Table 14Comparison of convergence time and final fitness.MethodConvergencetime (s)Final fitnessLatin hypercube2.19640.3615Sobol5.97840.3507Halton4.40850.3518Sierpinski3.87240.3552Proposed3.86360.3495 Table 14Comparison of convergence time and final fitness.MethodConvergencetime (s)Final fitnessLatin hypercube2.19640.3615Sobol5.97840.3507Halton4.40850.3518Sierpinski3.87240.3552Proposed3.86360.3495 Comparison of convergence time and final fitness. Computational time and efficiency analysis To evaluate the practical efficiency of the compared algorithms, we measured the average time required to segment a single image from the NASA dataset using an 8-level threshold. All experiments were conducted on the same hardware platform to ensure consistency. For a fair comparison focused on the search behaviour of the algorithms, execution time was recorded under the same experimental settings for all methods. The resulting average performance values and runtimes are reported in Table 15. Table 15Average performance and execution time comparison on the NASA dataset using 8-level thresholding.MethodPSNRSSIMFSIMQILVCORREPIMIFRuntime (s)AO33.17600.98220.88480.92830.95910.57942.43731.4033EO31.40210.96590.82430.87390.93690.50142.00340.8179PSO33.75330.86160.57970.64660.65090.24860.84751.0266WOA32.78280.95100.78560.83800.92490.46581.87160.7249Proposed 37.5360 0.9879 0.9259 0.9466 0.9795 0.7233 2.6157 1.4578Significant values are in [bold]. Table 15Average performance and execution time comparison on the NASA dataset using 8-level thresholding.MethodPSNRSSIMFSIMQILVCORREPIMIFRuntime (s)AO33.17600.98220.88480.92830.95910.57942.43731.4033EO31.40210.96590.82430.87390.93690.50142.00340.8179PSO33.75330.86160.57970.64660.65090.24860.84751.0266WOA32.78280.95100.78560.83800.92490.46581.87160.7249Proposed 37.5360 0.9879 0.9259 0.9466 0.9795 0.7233 2.6157 1.4578Significant values are in [bold]. Average performance and execution time comparison on the NASA dataset using 8-level thresholding. Significant values are in [bold]. As shown in Table 15, simpler baseline methods such as WOA and EO achieve lower execution times. The proposed AHA-GDA method requires slightly more time, with an average runtime of 1.4578 seconds. This modest increase in computational cost is mainly due to the channel wise energy curve computation, where spatial energy values are calculated separately for the red, green, and blue image channels. Although these components introduce a modest overhead, they play an important role in improving search space exploration and reducing the risk of being trapped in local minima. At the same time, the proposed method achieves the highest average values for PSNR, SSIM, FSIM, QILV, CORR, EPI, and MIF on the NASA dataset at 8-level thresholding. Therefore, the small increase in runtime is justified by the improvement in final segmentation performance. Overall, the proposed method provides a reasonable trade-off between computational time and segmentation accuracy for complex satellite image segmentation. Ablation study To support the inclusion of these three components, namely structured initialization, GDA refinement, and the energy curve, in the AHA framework, an ablation study was carried out. The ablation results for image 201080 at 8-level thresholding show that the complete model achieves the best performance across all evaluation metrics. Table 16 presents the quantitative results for test image 201080 when these components are omitted in a systematic manner. As observed from Table 16, each of the components plays an individual and unique role in ensuring segmentation accuracy. The exclusion of structured initialization leads to a sharp drop in the quality of signal reconstruction since PSNR drops from 39.0212 to 36.0021. On the other hand, the exclusion of the energy curve component leads to a decrease in the spatial and structural quality as revealed in SSIM of 0.9890 and QILV of 0.9640. The findings show that the energy curve is very critical in establishing context between the neighbors, which cannot be fully captured by the histogram. Finally, omitting GDA leads to lower FSIM value of 0.8779, indicating that GDA helps the solutions escape local optimum traps during later iterations.Table 16Ablation study results for image 201080.MethodImage nameSSIMFSIMPSNRQILVCORREPIMIFWithout initialization2010800.99440.881936.00210.98330.98720.89262.6450Without GDA2010800.99360.877937.34470.98180.98600.89212.6546Without energy curve2010800.98900.879737.56680.96400.98790.88842.8258With Initialization+GDA+Energy Curve2010800.99530.883839.02120.98580.98850.89312.9054 Table 16Ablation study results for image 201080.MethodImage nameSSIMFSIMPSNRQILVCORREPIMIFWithout initialization2010800.99440.881936.00210.98330.98720.89262.6450Without GDA2010800.99360.877937.34470.98180.98600.89212.6546Without energy curve2010800.98900.879737.56680.96400.98790.88842.8258With Initialization+GDA+Energy Curve2010800.99530.883839.02120.98580.98850.89312.9054 Ablation study results for image 201080. With Initialization+GDA +Energy Curve Statistical significance analysis In order to prove that the benefits in image segmentation provided by the proposed method are not due to the randomness of the metaheuristic technique but are mathematically sound, the non-parametric Friedman test was used. The analysis is carried out on the NASA dataset at the 8-level threshold. The segmentation results for all test images are evaluated using PSNR, SSIM, and FSIM measures. Table 17 demonstrates that, in all cases, the mean ranks indicate that the proposed method performs better than the competing methods. Also, the p-values for all three measures were found to be less than 0.001. Since these p-values are significantly smaller than the alpha value of 0.05, null hypothesis cannot be accepted with greater than 99.9% confidence.Table 17Friedman test-based statistical significance analysis.MetricProposedAOEOPSOWOAFriedman \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\chi ^2$$\end{document}p-valueSignificancePSNR1.803.233.852.733.4039.18< 0.001Highly significantSSIM1.282.103.154.953.53126.18< 0.001Highly significantFSIM1.032.083.204.983.73147.56< 0.001Highly significantLower mean rank values indicate superior performance, and the proposed method achieved the best rank across all evaluation metrics. Table 17Friedman test-based statistical significance analysis.MetricProposedAOEOPSOWOAFriedman \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\chi ^2$$\end{document}p-valueSignificancePSNR1.803.233.852.733.4039.18< 0.001Highly significantSSIM1.282.103.154.953.53126.18< 0.001Highly significantFSIM1.032.083.204.983.73147.56< 0.001Highly significantLower mean rank values indicate superior performance, and the proposed method achieved the best rank across all evaluation metrics. Friedman test-based statistical significance analysis. Lower mean rank values indicate superior performance, and the proposed method achieved the best rank across all evaluation metrics. Conclusion This paper has suggested a hybrid multilevel image segmentation model that is effective in integrating the methods of initialization, energy curve, the AHA and the GDA. Latin Hypercube, Sobol, Halton and Sierpinski sequences benefiting the initialization phase provided diversity of the population and improved the exploration of AHA through well dispersed candidate solutions. Also, structural consistency was maintained and the coherence of segmented regions enhanced because of the integration of spatial contextual information. The GDA was used as an adaptive acceptance mechanism, which further optimized the exploitation process by better accepting solutions and ensuring that the search did not get stuck in local minima. MCEM is employed as the objective function to effectively guide the threshold selection process and achieve superior fitness values. The experiments on benchmark color images determined that the suggested algorithm attained higher thresholding accuracy, and improved segmentation as seen on the generated visually consistent segmented images of the state of the art algorithms. Future extensions may include large scale applications such as medical diagnostics, hyperspectral image processing, and real time video segmentation, where energy curve representation could further enhance interpretability and decision making.