Contour Detection Model Based On Physiological Mechanism Of Visual Pathway

As one of the bases for mid-level and high-level vision tasks,contour detection is of great importance in the computer vision community.Deep learn-based codec contour detection methods usually adopt deep convolutional neural networks(VGG16,Res Net,etc.)as the encoding network,and the decoding network is designed by researchers.In the coding and decoding network,researchers pay more attention to the design of the decoding network,ignoring the influence of the coding network,which limits the performance of existing contour detection methods.Physiological studies have shown that the biological vision system is efficient and accurate in extracting contour features,and the initial design of the convolutional neural network is also inspired by the biological vision system.Although the existing contour detection methods of codec structure have some similarities with the biological visual system,they can not well reflect the physiological mechanism of the visual pathway and its processing process of visual information.Therefore,inspired by the neurophysiological mechanism in the biological visual system,this thesis studied the relationship between convolutional neural networks and visual pathways in-depth and proposed three bionic contour detection models based on convolutional neural networks by simulating the physiological mechanism in the retinal/LGN-V1-V2-V4-IT visual pathway,providing ideas for future contour detection research.The main research contents and innovations of this thesis are as follows:(1)In this thesis,the physiological mechanism of the retinal/LGN-V1-V2-V4-IT visual pathway is simulated,and a contour detection network combining a pre-enhancement network,coding network,and decoding network is proposed.Among them,the design of the pre-enhanced network simulates the information processing and transmission mechanism of retina /LGN and enhances the ability of the coding network to extract details and local features by combining it with the coding network that simulates the visual cortex region.Based on the hierarchical structure of the visual pathway and the function of integrating features in the IT layer,we design a novel decoding network and propose a down-sampling enhancement module in the decoding network,which enhances the feature integration capability of the decoding network.(2)In this thesis,the self-attention mechanism in Transformer is used to simulate the selective attention mechanism of V1,V2,and V4 in the ventral pathway to design the modulation network,and a new modulation and coding network is proposed by combining with the coding network,which effectively enhances the feature extraction capability of the coding network and realizes the selective extraction of global features.In addition,based on the function of the IT layer to integrate characteristic information,a new decoding network is designed.Different from the previous decoding network,it integrates top-down decoding and bottom-up decoding,uses the method of subsampling decoding to extract more adequate features,and then obtains better performance by integrating the features of upsampling decoding.(3)This thesis simulates the parallel processing mechanism of biological visual information and proposes a new bio-inspired lightweight contour detection network.The feature extraction network was designed by simulating the parallel pathway between ganglion cells,lateral geniculate body,and primary visual cortex region V1,which realized parallel processing and step-by-step extraction of input information,effectively extracted local features and detailed features in the image and improved the overall performance of the model.In addition,in the decoding network,a deep feature extraction module based on deep separable convolutions and residual connections is designed to integrate the output of the encoding network,which further improves the profile detection performance of the model.In this thesis,experimental evaluation is carried out on BSDS500,NYUD-v2,and BIPED datasets,and the effectiveness of the proposed method and module is verified.Experimental results show that the proposed model is highly competitive among existing contour detection models.