Improved DeepLabV3+remote sensing image segmentation algorithm based on attention mechanisms
DeepLabV3+has an efficient coding and decoding structure and is commonly used in image segmentation tasks.Aiming at the problems of imprecise segmentation target edges and hole defects in the semantic segmentation of DeepLabV3+high-resolution remote sensing images,an improved DeepLabV3+remote sensing image segmentation algorithm based on the attention mechanism is proposed.The ECBA(Efficient Convolutional Block Attention Module)attention mechanism is constructed,and ECBA is added to the DeepLabV3+backbone network Xception to enhance its feature extraction capability and obtain the attention-weighted high-level features.At the same time,ECBA is added to the encoder and decoder connection branch to get the attention-weighted low-level features.The decoder performs feature fusion of the two features to enhance the network's perception of the edges of different segmented targets as well as the interior of the same target.The experimental results show that the improved algorithm achieves a mean Intersection over Union(mIoU)and F1 index of 79.80％and 75.88％on the ISPRS Potsdam dataset,which is 11.06％and 6.32％better than the DeepLabV3+algorithm respectively.
万方数据
维普
