Research on semantic segmentation of high-resolution remote sensing images based on CW SwinUNet

The rapid development of artificial intelligence technology provides a new solution for semantic segmentation of high-resolution remote sensing images, but there are still challenges in capturing details and fusing global features in complex scenes. A CW SwinUNet semantic segmentation model that integrates convolutional neural networks and Swin Transformer is proposed. This model solves the problems of edge distortion and small target missed detection in high-resolution image segmentation by improving the encoder structure and combining deep convolution with SwinTransformer's self-attention mechanism. The experimental results show that the research model is superior in various indicators, with a mean improvement of 0.045 in mIoU, a pixel accuracy improvement of 0.037, and an F1 score improvement of 0.038. This model has improved the average intersection to union ratio index by 3.2% compared to SwinUNet and 8.7% compared to U-Net, indicating the high efficiency of the research team in image semantic segmentation. It can effectively support high-precision remote sensing image interpretation tasks and provide reference for the deep application of intelligent algorithms in the field of remote sensing image processing.