Semantic Segmentation of Remote sensing images via combining residuals and multi-scale modules

Abstract:
The residual network is designed for image classification, which is very effective for image classification. But for image segmentation, it is far from enough. Previous PSPNET uses multi-scale features and achieves great success. However, the previous PSPNET only used multi-scale features in the decoding part, so we integrated the advantages of residual network and multi-scale feature module and propose a residual multi-scale module. Experiments show that this module is very effective for semantic segmentation of high resolution images.