Shadow Detection of Remote Sensing Image by Fusion of Involution and Shunted Transformer

In the field of remote sensing image analysis, this paper addresses the challenges of accurately detecting edge details and small shadow regions through the introduction of STAISD-Net. This end-to-end network synergistically combines the capabilities of CNN and Transformer to enhance shadow detection accuracy. Our network architecture incorporates several innovations. In the encoder, we propose an improved multiscale asymmetric involution to extract detailed features across multiple scales. Additionally, we improve Shunted Transformer to extract global features, generating four-scale global feature information. In the decoder, we employ bilinear interpolation for upsampling and skip connections for fusing the high-level and low-level image features. Finally, we generate shadow masks by integrating feature maps of different scales. Comprehensive experiments conducted on the Aerial Imagery dataset for Shadow Detection (AISD) validate the effectiveness of STAISD-Net, showing that when compared to several state-of-the-art methods, our method demonstrates superior performance, achieving higher accuracy and the shadow detection results are more consistent with actual shadows.