Artificial Visual Network with Fully Modeled Retinal Direction-Selective Neural Pathway for Motion Direction Detection in Grayscale Scenes

The complexity and functional evolution of mammalian visual systems have always been a focal point in neuroscience and biological science research. The primary neurons that output motion direction signals have been a focal point of research in visual neuroscience for nearly 130 years. These neurons are widely present in the cortex and retina of mammals. Although the relevant pathways have been discovered and studied for almost 60 years due to experimental accessibility, research still remains at the cellular level. The specific functions and overall operational mechanisms of the component neurons in the motion direction-selective pathways are yet to be clearly elucidated. In this study, we modeled existing relevant neuroscience conclusions based on the symmetry and asymmetry of whole cells in the retina-to-cortex pathway and proposed a quantitative mechanism for motion direction selectivity pathways, called the Artificial Visual System (AVS). By tests based on 1 million instances of 2D, eight-direction grayscale moving objects, including 10 randomly shaped objects of various sizes, we confirm AVS’s high effectiveness on motion direction detecting. Furthermore, by comparing the AVS with two well-known convolutional neural networks, namely LeNet-5 and EfficientNetB0, we verify its efficiency, generalization, and noise resistance. Moreover, the analysis indicates that the AVS exhibits evident biomimetic characteristics and application advantages concerning hardware implementation, biological plausibility, interpretability, parameter count, and learning difficulty.