Abstract: Aimed at the problems of low segmentation efficiency and discontinuity of segmentation boundaries in the current image semantic segmentation, a semantic segmentation algorithm combined with multi-scale fusion enhancement and attention mechanism is proposed. The algorithm improved the original DeepLabv3 + network structure, proposed a feature extraction enhancement network structure in the encoder part, made full use of the feature information of each scale of the adjacent layer for fusion, and used the improved lightweight convolution attention module at the end of the decoder, making the object boundary segmentation more complete. Experimental verification on the Pascal VOC2007 and Cityscapes datasets shows that the accuracy of the proposed method is significantly improved compared with the original network.