Approaching human visual perception through AI-based representation of figure-ground segregation

Chakkai Yip¹Ezekiel Moroze¹Shigeaki Nishina²Arash Yazdanbakhsh^1*

• ¹Boston University, Boston, United States
• ²Kabushiki Kaisha Honda Research Institute Japan, Wako, Japan

Understanding how the visual system assigns borders to foreground objects is central to figure–ground perception, yet the computational principles underlying this process are still under investigation. In this work, we use deep convolutional neural networks (CNNs) as a controlled modeling framework to investigate how border-ownership (BOS) representations can emerge from hierarchical network processes. By training multiple network architectures on simple overlapping and occlusion-based stimuli and testing them under degrading contour conditions, we show that BOS can be inferred from feedforward computations. Analysis of fragmented borders revealed that networks rely disproportionately on junction-like configurations, highlighting the importance of geometric context over isolated edges. Network performance increased nearly linearly with the amount of context presented in the stimuli, supporting theories that horizontal and feedback connectivity supports BOS by reducing the visual information needed to establish a clear representation in deprived visual input conditions compared to exclusively feedforward architectures. Our visualization of internal representations further demonstrated a progression from local edge detection to spatially coherent BOS-specific representations along the CNN processing hierarchy. Together, these findings 2 clarify which aspects of BOS can arise from hierarchical processing and which likely depend on additional mechanisms present in biological vision.