Perceptual alignment in artificial vision: bio-inspired design and psychophysical evaluation

• Autores: Pablo Hernández Cámara
• Directores de la Tesis: Valero Laparra Pérez-Muelas (dir. tes.) , Jesús Malo (codir. tes.)
• Lectura: En la Universitat de València ( España ) en 2025
• Idioma: inglés
• Número de páginas: 220
• Tribunal Calificador de la Tesis: Maria Vanrell i Martorell (presid.) , Luis Gómez Chova (secret.) , Arash Akbarinia (voc.)
• Programa de doctorado: Programa de Doctorado en Ingeniería Electrónica por la Universitat de València (Estudi General)
• Calificación: Sobresaliente Cum Laude
• Enlaces
  • Tesis en acceso abierto en:  webges.uv.es (pdf)  RODERIC 
• Resumen

  • This thesis addresses the problem of perceptual alignment in computer vision: to what extent do deep neural networks perceive the visual world in a way comparable to humans? Although deep learning has revolutionized the field with state-of-the-art performance in segmentation, classification, and multimodal reasoning, its connection to the mechanisms and behaviors of human vision remains limited. This work analyzes the gap between biological plausibility, performance, and perceptual alignment, combining bio-inspired models with psychophysical evaluations. The first part focuses on biologically inspired computations. We integrated divisive normalization (DN), a canonical computation of the early visual cortex, into segmentation networks such as U-Net. Under adverse conditions (fog, low lighting, or low contrast), models with DN show greater robustness with only a minimal increase in parameters. However, when evaluated with the Decalogue¿a battery of psychophysical tests on contrast sensitivity and contextual masking¿they do not reproduce low-level human perception. Thus, biological plausibility improves robustness but does not induce human-like perceptual behaviors. The second part develops methodologies to measure alignment beyond accuracy. It introduces the Decalogue for low-level phenomena, a methodology to evaluate chromatic discrimination using MacAdam ellipses, contrast sensitivity function (CSF) tests in multimodal language models (MLLMs), and a framework to analyze abstraction levels in CLIP. Results show that alignment depends on data statistics and training objectives. Networks trained with richer chromatic distributions generate ellipses closer to those of humans, highlighting the role of the visual environment. CSF tests reveal that even advanced MLLMs fail to replicate basic human sensitivities. In CLIP, alignment varies across layers: at early stages, texture-based representations approximate human perception, but linguistic supervision shifts the model toward semantic abstractions, enhancing robustness but reducing low-level alignment. The third part examines the factors that determine alignment. Analyses of CNNs, Vision Transformers, CLIP, and MLLMs show that alignment is a multifactorial property emerging from the interaction between architecture, optimization objectives, data statistics, training duration, and reading strategies. Counterintuitively, increasing optimization or model size can reduce alignment, producing an inverted U-shaped relationship between accuracy and perceptual similarity. Linguistic supervision introduces a bias toward global shape representations at the expense of local details. Overall, alignment emerges more from task constraints (data and supervision) than from architectural changes. Conceptually, the thesis demonstrates that bio-inspired designs improve performance but do not guarantee human-like behaviors. Methodologically, it contributes systematic frameworks inspired by psychophysics to evaluate alignment in vision and multimodal models. Empirically, it clarifies how biological computations, chromatic environment, optimization, and language influence similarity with human perception. In conclusion, the thesis advances understanding of how artificial networks perceive and where they diverge from human vision. It shows that bridging performance and alignment requires going beyond architectural inspiration, toward data- and behavior-driven evaluations. These findings lay the groundwork for robust, biologically inspired architectures and standards that treat perceptual alignment, alongside accuracy, as a central objective in computer vision.