Semantic Segmentation of Remote Sensing Images Based on Swin Transformer

Abstract

With the evolution of Earth observation technology and remote sensing technologies, the amount of data available for high-resolution remote sensing images has exploded, and high-precision image segmentation has become a current research hotspot. Semantic segmentation technology is becoming increasingly important in fields such as urban planning, land use management, and autonomous driving. Large disparities within intraclass and modest differences intraclass are hallmarks of high resolution remote sensing pictures. Traditional image semantic segmentation methods rely on human-computer interaction and have poor generalization ability. When facing remote sensing images with rich types of ground objects and significant differences in target scales, the segmentation accuracy is not high. In this paper, we suggest an UperSwin decoder structure. The decoder includes several Swin transformer blocks and a fusion upsampling module, where the multi head contextual attention module in the Swin transformer block simultaneously uses multi-scale features and upsampling output features from the backbone network. In addition, the fusion upsampling module concatenates the backbone network features with the output features of the Swin transformer block, and then performs upsampling operations, preserving more detailed information. This article evaluates the accuracy and intersection ratio indicators on the Potsdam and Vaihingen datasets, verifying the feasibility and effectiveness of the model.