Abstract
Background: Mathematical models inspired by the Human Visual System (HVS) bridge the gap between biological perception and computer vision, by offering powerful frameworks for representing, enhancing, and analyzing digital images. From their initial development to contemporary applications, HVS models have demonstrated significant success in providing image-based solutions for complex real-world tasks. Objective: This work is an up-to-date systematic review of research on seven HVS models in real-world applications across six specific image processing tasks to evaluate, through five research questions, their contemporary relevance, adoption frequency, task-model preferences, and practical deployment domains. Methods: We systematically analyze seven prominent bio-inspired visual models across real-world applications: the Retinex Model (RM), Difference of Gaussians (DoG), Gabor Filters (GFs), Hermite and Steered Hermite Transforms (HT/SHT), the Steerable Pyramid (SP), Spiking Neural Networks (SNNs), and the Dual-Tree Complex Wavelet Transform (DT-CWT). Next, following the PRISMA guidelines, an initial literature search was performed in Scopus, yielding 595 publications published between 2020 and 2026, to assess the contemporary relevance of these visual models. From this initial search, 153 research works were selected based on the following inclusion criteria: (i) application to real-world image processing problems; (ii) coverage of at least one of six core image processing tasks: segmentation, watermarking/encryption, motion analysis and robotics, classification, pattern recognition, and enhancement; (iii) utilization of at least one of the seven mathematical visual models; (iv) document types restricted to peer-reviewed journal articles and book chapters; and (v) publications in the English language. Additionally, studies were excluded based on the following criteria: (i) focus on non-visual or non-image-based applications; (ii) utilization of alternative HVS models (e.g., Fourier log-polar mapping, Naka-Rushton photoreceptors, or Hopfield-based cortical networks); (iii) document types restricted to conference papers, editorials, and reviews; and (iv) publications written in languages other than English. Thus, through a systematic methodology, (a) the mathematical visual models were formally described by highlighting their biological foundations, analyzing their computational complexity and noise robustness, and outlining their recent developments in computer vision applications; (b) the selected publications were analyzed to show their statistical distribution across the seven models, publication modalities, and primary image processing tasks; and (c) the selected works were evaluated in terms of their distribution by model and year (via arithmetic mean, standard deviation, adjusted Fisher-Pearson skewness coefficient, and heatmaps), citation distribution, h-index, and mapping across real-world applications and specific vision tasks. Results: This study reveals that the mathematical visual models mimic distinct HVS stages and represent diverse receptive field structures. Additionally, this work finds that GFs, the HT/SHT, the SP, and the DT-CWT exhibit a high level of noise robustness, whereas GFs, the SP, and SNNS are the most elaborate bio-inspired visual models owing to their complex implementation steps and high parameter count. In contrast, the RM, DoG, and DT-CWT achieve the lowest computational complexity, but with limited image decomposition capabilities, while the HT/SHT offers competitive computational complexity alongside rich image decomposition capabilities. Conversely, RM and SNNs display the highest publication activity from 2020 to 2026 alongside notable temporal volatility, whereas GFs and DT-CWT maintain a stable yet dynamic trajectory across the same period. Moreover, studies employing DT-CWT, GFs, and SNNs exhibit sustained productivity and substantial domain impact, attaining h-index values of 10, 9, and 8, respectively. Regarding real-world applications, GFs represent the most widely deployed visual model (24.8%), followed by RM and SNNs with 20.3% and 19.6%, respectively. Task-specific analysis reveals distinct model specializations: RM leads in image enhancement (60%) and pattern recognition (33.3%) owing to its intrinsic color constancy properties, while GFs are predominantly selected for segmentation (44.4%) and classification (42.1%). Furthermore, DT-CWT and HT/SHT dominate watermarking and encryption applications (66.7%), whereas SP, SNNs, and RM serve as primary choices for motion analysis and robotics, representing 27.3% (SP) and 22.7% each (SNNs and RM) of task deployments. Regarding medical imaging, 33 (21.5%) of the 153 selected studies target diagnostic support or e-health applications. Within this domain, HT/SHT demonstrates the highest relative concentration, with 50% (5 out of 10) of its publications dedicated to medical applications. Conclusions: By systematically mapping the evolutionary trends and adoption metrics of these visual models across six primary vision tasks (segmentation, watermarking/encryption, motion analysis and robotics, classification, pattern recognition, and image enhancement) this work highlights current model specializations and application domains. Visual model selection depends on domain demands. Security utilizes transform-domain models (DT-CWT, HT/SHT); robotics requires spatiotemporal processing (SP, SNNs); segmentation and classification favors GFs and SNNs; and RM and GFs dominate enhancement and feature extraction. The resulting synthesis provides a clear taxonomy to guide future developments in bio-inspired computer vision. Finally, due to the wide diversity of implementation frameworks across the evaluated models and application domains, standardized empirical benchmarking of operational trade-offs and hardware metrics was beyond the scope of this review. These aspects represent promising avenues for future research, which will evaluate these technical parameters within dedicated, task-specific studies.