Research on Semantic Segmentation Algorithm Based on Enhanced DeeplabV3+

This paper proposes a semantic segmentation method based on an attention mechanism and multi-scale fusion for the semantic segmentation problem in computer vision. The multi-scale cascade module enhances the connection between each branch while maintaining the multi-scale hole convolution foundation. Since the deepening of the network causes the loss of some information, a self-attention module is introduced to fuse the attention through shallow and deep features, which better continues the features of the network. The efficacy of the proposed method is substantiated by experimental findings derived from the Cityscapes dataset, which demonstrate an enhancement in the model's accuracy.