Research On Ancient Mural Image Segmentation Method Based On Depth Separable Convolution Network

Chinese culture is broad and profound.As one of the main carriers of cultural recording and dissemination,ancient murals are important witnesses to the change of dynasties and history.However,through the changes of the times,under the dual influence of human and natural environment,these cultural treasures have been damaged to varying degrees,which seriously hinders the cultural relics researchers’ interpretation of the content of murals.In this case,image segmentation technology comes into the sight of researchers.Taking the traditional Chinese ancient murals as the research object,aiming at the problems of fuzzy target boundary and low efficiency of image segmentation in the process of ancient mural image segmentation,this paper proposes two multi category image segmentation models integrating lightweight neural network,The research methods and principles of the model are as follows:(1)Deeplab V3+ model integrating deep separable structure.Firstly,extracting lightweight features by hole convolution,and control the feature density of encoder according to the calculation resource budget.After that,the ASPP structure is used for multi-scale fusion of image feature information.The refined low-level features are combined with the output features after up sampling,which further enriches the feature information of mural image and realizes the research on the segmentation method of ancient mural image.(2)PSPNet model of fusion depth separable structure.Firstly,Mobile Net V2 with depth separable convolution network structure is used as the image feature extractor,and the point convolution is used to weight the image features in the depth direction.Then,the unique global pyramid module of pspnet network is used to extract the image feature semantic information after the maximum pool of feature map.Finally,the deconvolution method is used to up sample the low latitude features,and the feature maps of each level are spliced and fused to generate the prediction map.The main work of this paper includes the construction of mural image segmentation data set,the improvement of deeplabv3 + model and the construction of pspnet model integrating deep separable structure.Experiments verify the segmentation effect of two different image segmentation models on mural data set after integrating deep separable convolution network.The experimental results show that the improved model improves the accuracy and efficiency of image segmentation in varying degrees.