【摘要】 针对无人机航拍自运动视觉场景中运动视差导致传统方法精度差问题，基于蝗虫视觉神经结构特性，借助生物视信号中心环绕机制、视觉短期记忆机理，提出一种生物启发的人工视觉系统——自运动前景提取神经网络（SFENN）。SFENN由5个复眼神经层（R、L、M、Lo、LP）构成。其中，R层接收自运动视觉信号；L、M层采集全局视野域中的自运动角点特征与轮廓信息；Lo与LP层提取前景目标，对其降噪处理并向外输出表征前景轮廓视信息的膜电位兴奋。实验研究结果表明，SFENN在无人机航拍的自运动视觉场景中能有效提取前景目标对象，与SOTA模型相比体现更佳检测准确率和鲁棒性。更多还原

【Abstract】 Aiming at the problem of poor target extraction accuracy of traditional methods due to motion parallax in self-motion vision scenes of unmanned aerial vehicle(UAV) aerial photography, a bio-inspired artificial vision system named self-motion foreground extraction neural network(SFENN) is proposed. This system is based on the neural structure properties of the locust compound eye, the center-surround visual signal mechanism and the visual short-term memory mechanism found in biological vision systems. The proposed neural network consists of five neural layers, including R, L, M, Lo, and LP layers. Wherein the R layer receives self-motion visual signals; L and M layers collect the self-motion corner features and contour information in the whole field of view; Lo and LP layers extract the foreground targets, process them with noise reduction and output membrane potential excitation, which represents the visual information of the foreground contours. Experimental studies demonstrate that SFENN can effectively extract foreground targets in self-motion visual scenes of UAV aerial photography and exhibit better detection accuracy and robustness performances compared with the SOTA models.更多还原