STRSNet: A Spatial Transformer-based Remote-sensing Segmentation Network

Multi-scale feature extraction is commonly adopted in high-resolution remote-sensing semantic segmentation tasks, in which images are reshaped into multiple sizes to generate features from different levels. However, larger size images contain less detailed information and smaller images lose more contextual information, and both of them restrict the performance of remote-sensing semantic segmentation neural networks. In this paper, a Spatial Transformer-based Network for Remote-sensing Segmentation(STRSNet) is proposed to deal with this problem. Local images of smaller sizes are cut from the center of larger images, and both of them are input to STRSNet separately. Local features from cropped local images and contextual features from larger images are extracted respectively in STRSNet to learn both detailed and contextual information and capture more long range dependencies, and a Spatial Enhancement and Perception Transformer(SEPT) is used for feature fusion. SEPT contains a Polar Coordinates Spatial Enhancement module(PCSE) and a Spatial Multi-layer Perceptron module(SMLP) to build fine-grained local spatial information. Experimental results on Land-cover Dataset with Gaofen-2 satellite(GID) have shown that the proposed STRSNet has outperformed the compared methods.