A Full-Scale Feature Extraction Network for Semantic Segmentation of Remote Sensing Images

Abstract

Semantic segmentation of high-resolution remote sensing images is an important and challenging task. Aiming at the difficulty of accurate segmentation of confusing objects in high-resolution remote sensing images, we propose a full-scale feature extraction network (FSFN et). The proposed architecture follows the encoder-decoder paradigm. The encoder adopts ResNeXt50 to fully extract features, and inputs the four-scale output of the encoder to the global feature extraction module (GFM) as part of the decoder input. The decoder uses group convolution residual module combined with attention mechanism (GRM) to refine features, and employs dense connection (DC) to strengthen feature propagation. In addition, the third and fourth decoding blocks are added with deep supervision (DS) to speed up the convergence of the network. We evaluate our proposed architecture on the Vaihingen and Potsdam datasets provided by the International Society for Photogrammetry and Remote Sensing (ISPRS). The experimental results show that the proposed approach can significantly improve the segmentation performance, reaching 89.8% and 89.3% overall accuracy (OA) on the two datasets, respectively.