| The retinal physiological functions and human visual perception mechanism have always been research focuses in the fields of neurobiology and signal processing. The aim of the researches is to establish the corresponding mathematical models for application based on revealing and analyzing the neural computing mechanism. Recently, with the rapid development of computer technology and the latest neurophysiological findings, more and more attentions are drawn to the applications to machine vision of the neural models that are constructed based on retinal computing mechanism.As an important feature in an image, an edge is the boundary between the areas with different textures or intensity discontinuity. It is significant to reach an acceptable balance between removing noise and locating edges for image segmentation and recognition as well as other advanced fields of image processing technology. According to the basic mechanism by which retina processes information in a parallel-cross way through the channels for respectively processing information carried by luminance, colors and motion, and on summerizing the retina functions and its neural mechanisms respectively on edge detection or motion computation, Retina computational mechanism networks for luminance edge processing 、 motion detection, or parallel-cross edge enhanced extraction are proposed in this paper. The main contributions of this paper include:(1) On summerizing the retinal physiological mechanisms for detecting edges, a neural model of a retinal luminance channel is proposed. It includes the topology of the network, the mathematical input-output equation as well as the receptive field properties for neurons in each layer, etc.;(2) On studying systematically the retinal physiological mechanisms of computing object motion, a neural model of a retinal motion channel is proposed. It includes the direction selectivity mechanism with DSRGC, the input-output relationship and the receptive field properties with neurons in each layers of the motion channel, and the topology of the neural network, etc.;(3) On synthesizing the characteristics of the luminance channel and the motion one, a hybrid neural network, which is constructed by a luminance channel and a motion one, for detecting edges is proposed.(4) Analysis and comparison of several performance evaluation methods are implemented. Emphases are put on the complexity of the algorithm and the adaptability of the method as well as some related comprehensive analysis and comparison, which guaranteeing a reasonable and accurate assessment of edge detection quality;(5)Simulation experiments are implemented and experimental results are compared and analysized. Experimental results show that, the retina Net-Edge algorithm shows a better performance on edge detection, higher edge location, lower miss or false rate, not only the image edges are detected effectively and accurately, but also the important texture details are reserved. The overall performance is much more superior to the traditional gradient-based edge detection algorithm as well the latest neural network based tremble. The moving object can be well detected under the retina Net-Motion algorithm, not only the target spot, but also the derection. The retina Net-Edge Enh algorithm improves the edge detection performance of the retina Net-Edge algorithm, to sharpen edge on the premise of good pixel-accuracy, and to effectively weaken or wipe off the false edge.The works on neural model for edge detecting, provide us with new ways to enhance the clarity of the image edges and to reduce noise. All of the research achievements can be applied to various applications, such as target tracking, object analysis, missile guidance, medical diagnosis and image matching based guidance, etc., for achieving the goals of real-time processing, high precision, and less power requirements. Besides, it also help people deepen their understanding of the principle of visual neural computing, which means an important theoretical significance. |
PDF Full Text Request