Abstract: Cognitive neuroscience studies have found that there are direction differential neurons (OMS-DS) in the vertebrate retina, which have the visual neural response property of local-global direction difference preference, which helps to distinguish the motion difference between foreground localization and background globalization in a dynamic visual scene; however, there has not been any report of computational model for the study of this neural property in visual motion perception problem. To address this problem, a directional differential motion perception neural network (dirDMPNN) is proposed based on the salamander retinal differential motion response properties. The proposed neural network contains both presynaptic and postsynaptic parts. The presynaptic network perceives low-order visual cues triggered by motion changes in the visual field domain; the postsynaptic network realizes the response output based on the gap-junction coupling mechanism of overlapping receptive field to deferentially discriminate foreground and background directions. Systematic experimental studies show that the dirDMPNN perceives foreground-background direction motion cues from translation self-motion in the visual field domain and outputs strong neural spike responses to its differential motion patterns. This study is inspired by the neural mechanisms of the biological visual brain and focuses on visual information processing. It offers novel insights and methods for addressing motion perception, recognition, target detection, and tracking in self-motion visual scenarios, such as those encountered in autonomous environments including spacecraft, autonomous driving, and robotic navigation.

Key words: directional differential perception, gap junction coupling, overlapping receptive fields, directional differential neuron, retinal nerve, visual motion perception, self-motion visual scene

摘要： 认知神经科学研究发现，脊椎动物视网膜存在方向差分神经元（Object Motion Selective and Direction Selective, OMS-DS），具有局部-全局方向差异偏好的视觉神经响应特性，有助动态视觉场景中区分前景局部和背景全局之间的运动差异，但目前尚未有该神经特性在视觉运动感知问题研究的计算模型报道。针对该问题，基于蝾螈视网膜差分运动响应特性，提出一种方向差分运动感知神经网络（Directional Differential Motion Perception Neural Network, dirDMPNN）。所提出的神经网络包含突触前和突触后两部分。突触前网络感知运动变化在视野域中引发的低阶视觉线索；突触后网络基于重叠感受野间隙连接耦合机制以对前景、背景方向差分实现响应输出。系统性实验研究表明，dirDMPNN能感知视野域中平移自运动的前景-背景方向运动线索，并对其差分运动模式输出强烈神经尖峰响应。论文工作涉及生物视脑神经机制启发的视觉信息加工处理，可为自运动视觉场景，诸如空间飞行器、无人驾驶、机器人自主导航等自治环境下的运动感知与识别、目标检测与跟踪问题的研究与解决提供新思想、新方法。

关键词: 方向差分感知, 间隙连接耦合, 重叠感受野, 方向差分神经元, 视网膜神经, 视觉运动感知, 自运动视觉场景