Abstract: The existing semantic segmentation algorithms are difficult to deploy on mobile devices due to the complex model and a large amount of computation.A new semantic segmentation algorithm based on hybrid attention is proposed.This algorithm is an asymmetric encoder-decoder structure.The encoder part combines depth-wise separable convolution anddilated convolution to design an efficient residual module to extract image features at different levels of the network.It pays more attention to spatial position information in the shallow layer and enhances semantic information extraction in the deep layer.In the decoder part,a hybrid attention feature fusion module is designed,which uses spatial attention to strengthen the spatial location information in the shallow layer and channel attention to enhance the expression ability of key information in the deep feature map.It can effectively integrate the spatial information and context information in the feature map of different levels,strengthen the expression of semantic information,and reduce the loss of image information in the fusion process.Finally,the segmentation results are predicted by using the classifier.A large number of experiments show that the proposed algorithm achieves 93.2% PA and 73.2% mIoU in Cityscapes,respectively,and achieves 38FPS with 1.62×10⁶ reference on Tesla V100 GPU.In Pascal VOC 2012 data set,PA and mIoU reaches 92.4% and 74.8% respectively.Experimental results show that this algorithm can effectively and quickly complete the task of city scene image segmentation.

Key words: Deep learning, Semantic segmentation, Real-time, Feature fusion, Attention mechanism