CrossFormer Embedding DeepLabv3+ for Remote Sensing Images Semantic Segmentation

High-resolution remote sensing image segmentation is a challenging task. In urban remote sensing, the presence of occlusions and shadows often results in blurred or invisible object boundaries, thereby increasing the difficulty of segmentation. In this paper, an improved network with a cross-region self-attention mechanism for multi-scale features based on DeepLabv3+ is designed to address the difficulties of small object segmentation and blurred target edge segmentation. First, we use CrossFormer as the backbone feature extraction network to achieve the interaction between large- and small-scale features, and establish self-attention associations between features at both large and small scales to capture global contextual feature information. Next, an improved atrous spatial pyramid pooling module is introduced to establish multi-scale feature maps with large- and small-scale feature associations, and attention vectors are added in the channel direction to enable adaptive adjustment of multi-scale channel features. The proposed network model is validated using the Potsdam and Vaihingen datasets. The experimental results show that, compared with existing techniques, the network model designed in this paper can extract and fuse multi-scale information, more clearly extract edge information and small-scale information, and segment boundaries more smoothly. Experimental results on public datasets demonstrate the superiority of our method compared with several state-of-the-art networks.