SAR remote sensing image segmentation algorithm based on improved DeepLabV3+

Due to the shape, size and diversity of water bodies in SAR remote sensing images as well as the complexity of the scene, the problem of loss of fine water bodies often occurs in the water body segmentation task, which leads to the low accuracy of water body segmentation. Therefore, in this paper, we improve on the DeeplabV3+ network. The network first uses the lightweight MobileNetV2 backbone network to reduce the number of parameters and computation of the net-work and to increase the model training speed. Secondly, in the encoder phase, the dense empty space pyramid pooling (DenseASPP) module is used to obtain denser feature pyramids and larger sensory fields, and bar pooling branches are added to enable the network to capture long-distance dependencies more efficiently and to improve the segmentation accuracy of the network for water bodies. Finally, in the decoder stage, the Attention Gate module is introduced before the fusion of low-level features and high-level features to suppress irrelevant information in the image and highlight im-portant features of the target. The improved network is named DSA-DeepLabV3+ network. The results of experiments performed on the home-made dataset show that the segmentation accuracy of DSA-DeepLabV3+ is better than DeepLabV3+ and its comparison methods.