Research On Method Of Image Semantic Segmentation Based On Deep Learning

In the research of Image Semantic Segmentation based on Deep Learning(ISSbDL),the Image Semantic Segmentation(ISS)method based on the encoder-decoder model has problems such as loss of pixel spatial position information and inefficient use of image context.In this paper,two improvements are made to the encoder-decoder model and based on these two improvements,an ISS method based on improved encoder-decoder model is designed.The main contributions of this paper are as follows:(1)The encoder-decoder model is improved:The encoder-decoder model is improved,and the basic network structure of the encoder module is redesigned to enhance the feature extraction capability of the encoder.On the one hand,multiple atrous convolutions are connected in a more densely connected manner,and the primary feature extractor is modified to present a Densely Connected Atrous Convolution Network(DenseAtrousCNet).On the other hand,several atrous convolutions with different dilation ratios are combined in a dense connection idea,and the Atrous Spatial Pyramid Pooling is improved for feature fusion,and a Densely Connected Global Atrous Spatial Pyramid Pooling(DenseGlobalASPP)model is proposed.DenseAtrousCNet and DenseGobalASPP are used in the encoder module for feature extraction,and deconvolution is used in the decoder module for feature recovery.(2)An ISS algorithm based on improved encoder-decoder model is proposed:Based on the above two improvements,the traditional encoder-decoder model is optimized and the framework structure of the model is redesigned.The traditional encoder-decoder model is optimized,and a Densely Connected Atrous Spatial Pyramid Pooling Deconvlution Network(DenseASPPDeconvNet)is proposed.Experiments in the PASCAL VOC 2012 public dataset show that DenseASPPDeconvNet captures more dense features and image contexts and improves ISS accuracy.