一种无人机航拍图像多尺度目标语义分割方法OA

To address the limitations of small target segmentation accuracy and cross-scale feature alignment in multi-scale target semantic segmentation tasks for UAV aerial images,this paper proposes the TransSeg model.The model integrates the global modeling capability of Swin Transformer with the symmetric encoder-decoder architecture of U-Net,and introduces D-ASPP and residual attention mechanisms.The loss function employs a collaborative optimization strategy combining Dice Loss and Focal Loss.Experiments were conducted on the UAVid dataset,with comparisons made against mainstream models including DANet,BiSeNet,and SegFormer,validating the effectiveness of TransSeg in segmenting 8 types of urban scene targets,and demonstrating strong performance on a custom dataset.The results show that the model achieves an average IoU of 66.9％,an improvement of 0.9％over SegFormer,with IoU increases of 4.1％and 5.6％for small targets such as moving vehicles and pedestrians,respectively.The research highlights the critical role of global context modeling and dynamic multi-scale fusion mechanisms in enhancing segmentation performance,providing a new solution for UAV remote sensing image processing.