Abstract: In the current semantic segmentation convolutional network,the spatial and detail information is gradually lost with the deepening of the convolutional layer,resulting in inaccurate segmentation of boundary parts and small objects.Meanwhile,the local feature capability of convolution restricts the network's ability to obtain effective global modeling,resulting in confusion of internal segmentation of objects.Aiming at these problems,a multi-path semantic segmentation algorithm based on edge optimization and global modeling is designed.The algorithm proposes a multi-path adjacent dislocation fusion network.Four branches of different resolutions are interlaced and fused adjacently.In order to reduce the loss of spatial information and detail information,the detail information between the adjacent four different resolution paths is fused,and the semantic information is fused between the tail of the high-resolution path and the header of the low-resolution path.The adaptive edge feature module is proposed to obtain edge features which are integrated into the middle layer and depth supervision layer of the network to enhance the expressive ability of edge features and the segmentation effect of small objects.The Transformer global feature module is proposed,which uses different convolutions for downsampling operations to reduce the length of self-attention sequences and fuse channel information and self-attention information to obtain effective high-level semantic global information.Experimental results show that the mIoU value on the CamVid test set reaches 76.2%,and the mIoU value on the Cityscapes validation set reaches 79.1%.

Key words: Semantic segmentation, Multi-path, Edge optimization, Deep supervision, Global modeling