FML-Swin: An Improved Swin Transformer Segmentor for Remote Sensing Images

Semantic segmentation of urban remote sensing images is a very challenging task. Due to the complex background, occlusion overlap and small scale target of urban remote sensing image, the semantic segmentation results have some defects such as target confusion and similarity, target boundary ambiguity, and small scale target omission. To solve the above problems, a feature-interactive fusion and multi-scale detail sensing lightweight enhanced Swin Transformer (FML-Swin) is proposed. The model includes several key components: feature interactive fusion transformer (FIFT) module, which enhances the model’s focus on current channel features; multi-scale detail sensing (MSDS) module, specifically designed to capture small scale features and details in remote sensing images; and lightweight enhanced squeeze excitation (LESE) module, which enriches the semantic feature information contained in the input image while maintaining a lightweight design. With limited training rounds, the model achieves a mIoU accuracy of 78.58 on the multi-class semantic segmentation task of the Potsdam dataset, exceeding SegNeXt 0.49. In addition, on the multi-class semantic segmentation task of the Vaihingen dataset, the mIoU accuracy of the model is 74.75, which is higher than SegNeXt 0.17. These results demonstrate the validity of the model.