Biological Visual Computing Model And Application

At present,image processing based on biological vision bionics has very broad research prospects and value.It is very challenging to explore the application of biological vision systems in image feature extraction and other basic aspects;through in-depth study of the internal mechanism,and the establishment of a similar vision system The mathematical model of the bionic vision calculation of the working mechanism provides a new theoretical guidance for the subsequent bionic vision system in the actual image processing research and application,and it is applied to the fields of contour detection,image segmentation and other fields,and very good experimental results have been achieved.First,the physiological processing mechanism of the pulse information flow between the levels of the visual cortex on the primary visual pathway is studied,and a visual neural computing model based on the frequency-divided visual mechanism is proposed,and it is applied to the extraction of prominent contour features of the image.Experimental results;secondly,a visual network computing model based on multi-level feature channel optimization coding is proposed,and a multi-level feature extraction visual neural computing model is established for image preprocessing.After an improved full convolutional neural network deep training,the image can be realized efficiently and accurately Contour detection;Finally,with retinal vessel segmentation as a specific application,the visual network computing model used for contour detection is transferred to the segmentation task,and a multi-scale dual-path convolutional network method for retinal vessel segmentation is proposed.The main research work and results of this paper are as follows:(1)Establish a calculation model based on the frequency-divided vision mechanism and apply it to the extraction of significant contour features of the image.To simulate the frequency division characteristics in the process of visual information flow transmission,firstly construct the frequency domain response characteristics of the LGN receptive field,use the Gaussian derivative function to simulate the frequency division characteristics of the LGN classical receptive field for visual information;secondly,according to the spatial frequency and orientation tuning It also introduces contrast features to construct a receptive field that is oriented sensitive and contrast-adapted.Through the detection of the difference between the center of the visual receptive field and the peripheral information,the selective suppression of multiple features is achieved;finally,the frequency division of the primary visual cortex is used The visual information flow fusion mechanism realizes the parallel processing and relevance transmission of the visual information flow to achieve the effective adjustment and complete fusion of the extraction of significant contour features.(2)Establish a calculation model based on multi-level feature channel optimization coding,and apply it to the contour detection of natural images.In order to solve the generalization of feature information in the traditional convolutional neural network detection process and the problem of heavy dependence on the number and quality of samples,the Gabor filter is first introduced to simulate the multi-level feature response on the visual path;secondly,the frequency domain separation is achieved through NSCT transformation,Get the pre-processed samples of the network;finally build an improved fully convolutional neural network model for optimized coding to achieve image contour detection.(3)A calculation model based on multi-scale dual-path convolutional network is established and applied to the segmentation of retinal blood vessels in medical images.Focus on the analysis of small blood vessels and low-contrast areas in retinal images,and simulate the multi-path transmission mechanism of visual information flow.First,according to the difference in sensitivity of Gabor filters to blood vessels of different widths at different scales,the roughness of the overall characteristics is obtained.Vessel feature maps and small blood vessel feature maps that retain local details;then,according to the characteristics of thick and thin blood vessel feature maps,a dual-path deep convolutional network is constructed,in which the overall blood vessel segmentation network achieves end-to-end fast segmentation;the local blood vessel segmentation network achieves fine The accuracy of blood vessel segmentation is improved;finally,the dual-path blood vessel segmentation image is logically ORed to realize the supplementary information of the small blood vessels that are easy to be ignored in the overall blood vessel segmentation process.