Pyramid Pooling And Spatial Attention Optimized Deep Semantic Segmentation

As a pixel-level image classification task,semantic segmentation algorithm classifies pixels judged to the same semantic label into the same category,which is one of the research hotspots in the field of computer vision.Nowadays,this technology has been widely used in many important fields such as medical image analysis,vehicle autonomous driving and geographic information system.The thesis further improves the performance of semantic segmentation method based on deep neural network by feature optimization.The realization of semantic segmentation needs to consider two factors,namely intra-class inconsistency and the inter-class indistinction.Therefore,it is of great research value to design a network framework that can extract effective features to further obtain accurate segmentation results.Firstly,this thesis introduces the research background and significance,present situation and development trends of semantic segmentation algorithm at home and abroad,and then further introduces the basic principle of semantic segmentation,then the problem of object edge and image detail blurring caused by the lack of spatial features in the traditional discriminative feature network for semantic segmentation is proposed,and puts forward deep semantic segmentation algorithm based on spatial attention(SA)optimization.Finally,to further aggregate different receptive field of context information,pyramid pooling and spatial attention(PPSA)optimized deep semantic segmentation is proposed in this paper.To obtain more effective spatial information,this thesis proposes a deep semantic segmentation algorithm based on SAM optimization.By introducing spatial attention module to aggregate space characteristics of low stage in the network and further select the discriminative features for remedying the loss of spatial information due to the continuous convolutions and pooling operations.Moreover,we adopt channel attention modules to obtain the context information generated from the high stages of the network.The performance of the proposed method was evaluated by PASCAL VOC 2012 and Cityscapes databases,and the effectiveness of the proposed method was verified by the final experimental results.In order to obtain more context information,this thesis proposes a deep semantic segmentation algorithm based on PPSA optimization.A pyramid pooling module is added at the end of the network to aggregate the context information of different receptive fields.Moreover,the introduced spatial attention module and channel attention module are used to obtain more effective spatial and context features.Finally,Pascal VOC 2012 and Cityscapes database were used to test the performance of the method.Experimental results show that the proposed method can achieve an accurate semantic segmentation result.In conclusion,this thesis proposes optimization algorithm from two perspectives of obtaining more effective spatial information and context information respectively.The proposed deep semantic segmentation algorithm based on PPSA optimization can solve the problems of intra-class inconsistency and inter-class indistinction,thus further improving semantic segmentation results.