Brain-like Computing For Visual Perception And Its Application

With the integration of brain information processing and multitasking collaboration,attention,and memory mechanisms,the visual system has superb information processing capabilities.Inspired by this,this paper attempts to model its simulation and apply it to image processing by exploring the information processing method of visual cortex.Firstly,considering the fine vision of the ventral visual pathway and the brain mechanism of local and global visual perception,the Contour Perception Based on Multi-path Convolution Neural Network is constructed from the perspective of local and global separation.Secondly,the primary visual cortex and advanced are simulated.Based on the information processing characteristics between the visual cortex,the Image-Dehazing Method Based on the Fusion Coding of Contours and Colors is constructed.Finally,the hierarchical feature coding model is applied to image dehazing and few sample recognition tasks.The main research contents of this topic are as follows:(1)A new contour sensing method for multi-path convolutional neural networks is proposed.Firstly,the Gaussian pyramid scale decomposition is used to obtain the low-resolution sub-graphs that represent the visual overall information.The boundary response sub-graphs describing the detailed features are obtained by the 2-dimensional Gaussian derivative function.Secondly,the sub-network with sparse coding characteristics(Sparse-Net)is constructed.Realize the rapid detection of the overall contour,construct a sub-network with redundant enhancement coding characteristics(Redundancy-Net)to achieve local detail feature extraction;finally,realize the fusion of the overall perception and local detection of the contour response through the pixel contrast relationship.Refined perception of the contours.Taking the BSDS500 library as the experimental object,in the GTX1080 Ti environment,the detection speed of the overall contour by Sparse-Net alone reaches 42 frames/s;the data of the detection index data set by Sparse-Net and Redundancy-Net is optimal(ODS).The optimal image scale(OIS)and average accuracy(AP)were0.806,0.824,and 0.846,respectively,which were superior to other comparison methods.(2)A new image dehazing method based on fusion of contour features and color features is proposed.Firstly,the contour feature extractor is constructed to extract the contour features of the image.Secondly,the low-level feature encoder is constructed to extract the low-level features and merge into the contour features.Then,the advanced semantic encoder is constructed to realize the deep analysis of the semantic information of the back-propagation process.Finally,The image is dehazed by combining the output of low-level feature coding and high-level semantic coding.Thesynthetic and natural foggy images were selected as experimental objects.The peak signal-to-noise ratio(PSNR/d B)and structural similarity(SSIM)of the composite images were improved by45.13% and 1.91%,respectively,compared with the dark channel prior(DCP);in the natural image,the average gradient(AG)and information entropy(EY)indicators were 7.22% and 12.70% higher than DCP,respectively.Compared with other dehazing effects,the proposed algorithm has strong robustness and improved color distortion and halo.(3)Apply low-level feature coding and advanced semantic coding models to few sample recognition.Firstly,the feature extraction is performed by using the hierarchical coding model.Then,the cosine distance measurement module is used to calculate the literacy of the test sample and the support sample.Finally,the information redundancy of the cosine similarity matrix is filtered by the Top K mechanism.Through the visualization method,the specific role of each convolution kernel in the recognition task of the model is presented,which facilitates a clearer understanding of the recognition process.In the Mini Image Net data sets and the previous algorithm,the recognition rates of 5-way 1-shot,5-shot and 10-shot are 60.04%±0.97%,76.92%±0.55%,and 79.19%±0.33%,respectively,are superior to other algorithms.