Multiscale Feature Extraction Fusion Network for Semantic Segmentation of High-Resolution Remote Sensing Images

In recent years, semantic segmentation of high-resolution remote sensing images has attracted much attention. However, the large size difference, complex background, and numerous small objects in remote sensing images bring challenges to the existing methods. For this purpose, we developed a high-resolution remote sensing image semantic segmentation network based on a multi-scale deformable Convolutional module (MDC) and multi-scale Attention Fusion module (MFF). Firstly, a multi-scale deformable convolution module is used to extract features of different scales and shapes. Then the multi-scale attention fusion module is embedded into the FPN module, and the multi-scale feature fusion is realized better through the mutual guidance of deep feature and shallow feature. In the end, we achieved 83.40% average cross-linking (MIoU) on the Postam dataset and 74.97% average cross-linking (MIoU) on the Vaihingen dataset.