Decoding Illusion Perception: A Comparative Analysis of Deep Neural Networks in the Müller-Lyer Illusion

Visual illusions, as powerful tools for revealing the strategies and limitations of brain visual processing, have garnered attention in neural network research. This study aims to investigate the performance of various deep neural networks (DNNs) in handling the Müller-Lyer illusion task, thereby providing insights into the mechanisms of the brain's visual system in processing visual illusions. We adopted two types of visual stimuli and recruited 12 participants for a length-matching task. Based on the participants' perceptual data, we employed multiple classical neural network models to test their responses to each type of stimuli. To analyze the differences in model performance, we employed representational dissimilarity matrices (RDM) analysis and utilized Grad-CAM techniques to visualize neural network activation maps. The results revealed that some neural networks could simulate human visual illusion phenomena but exhibit differences in representational similarity and feature preferences. These findings provide a preliminary understanding of the computational principles and mechanisms of the brain's visual system and offer insights for future research in building more robust and efficient neural network models.