Research On Image Semantic Segmentation Method Based On Deep Learning

With the continuous improvement of the theory related to deep learning and the continuous updating of hardware devices,computer vision tasks based on deep learning have become one of the hottest research projects.Among them,the most challenging topic is the image semantic segmentation technology.Image semantic segmentation is a way which understands and recognizes picture content in pixel level,and segment each pixel by semantic information.The quality of semantic segmentation results affect other computer vision tasks directly,such as image classification.scene analysis,and target detection and so on.Image semantic segmentation technology is used in the fields of automatic driving,medical image analysis,and smart home currently.Therefore,the research of image semantic segmentation technology has vital practical significance.Classical semantic segmentation models are generally constructed using convolution neural networks.The center of such models is to improve the semantic segmentation accuracy by improving or optimizing the network structure.From this point of view,this paper proposes a reasonable improvement strategy based on the shortcomings of existing semantic segmentation models.so that the model can extract more effective semantic information and improve segmentation accuracy.The detailed work is as follows:(1)This paper proposes an image semantic segmentation model based on multi-scale feature fusion and hybrid dilated convolution.First,a single branch network DCNN is constructed by the cascaded deep residual network,the hybrid dilated convolution,and an improved atrous spatial pyramid pooling module.The cascaded depth residual network captures long-span local context features by increasing the network depth among DCNN.Introducing a hybrid dilated convolution in a cascaded depth residual network can effectively mitigate the "grid" effect.Secondly,the input image is scaled to 4 different scales,and they are input a the parallel network containing 4 DCNN branches to extract the multi-scale features of the object.Finally,the output of each branch of the parallel network is obtained through the multi-scale feature fusion layer,and the fusion result is obtained.The fusion result is iteratively optimized with the fully connected conditional random field,and the fine segmentation result is obtained.Experiments show that the proposed semantic segmentation model has strong feature expression ability and can effectively improve the semantic segmentation precision.(2)In order to model the global context pixel dependence,provide relevant global feature information.The recurrent layer constructed by the bidirectional GRUs is stacked on the convolution neural network and an image semantic segmentation model based on deep residual network and bidirectional GRUs is proposed.Firstly,the local feature is extracted from the input image by the depth residual network,and three coordinate channel layers are added to feature map.Then,the new feature map is sent to the recurrent layer to model the context pixel dependence and adding three coordinate channel to the output feature map to form a new feature map and sending it to another recurrent layer.Finally,the output of the other recurrent layer is sent to the atrous spatial pyramid pooling module to extract the multi-scale context information of the image,and get a rough score map.The upsampling operation is performed on the score map to output a semantic segmentation result.Experiments show that the local feature information extracted by CNN and the correlation between image pixels modeled by bidirectional GRUs can effectively model local and global context features.Adding coordinate channel layer to feature map can enrich the coordinate information of the model.Improve the generalization ability of the model and produce semantic segmentation results with high resolution and precise boundaries.