When using Swin Transformer to study remote sensing images, a pre-trained model is used. Although the use of pre-trained models will accelerate convergence, it also causes the model to be unable to explicitly create semantic context information. To alleviate this problem, we designed a semantic attention layer based on a gated residual structure in this article and inserted it into the Swin Transformer network architecture. This module can build pixel-level semantic associations, enhance relevant semantic information, and suppress irrelevant information. We conducted experiments on the LoveDA and Potsdam datasets, respectively. The Miou value on the LoveDA dataset is 54.57%, an increase of 0.91% compared to the original model, and the Miou value on the Potsdam dataset is 79.11%, an increase of 0.63% compared to the original model.
Kong et al. (Fri,) studied this question.