关键词: 无人机正射影像, 语义分割, U-Net3+, ResNet50, 卷积注意力模块

Abstract: To address the limitations of insufficient feature abstraction and cross-scale feature redundancy in semantic segmentation of unmanned aerial vehicle (UAV)orthophoto images using the U-Net3+ model,this study proposes an improved U-Net3+ architecture.The improvement incorporates ResNet50,a deep convolutional neural network based on residual network,as the backbone for feature extraction.Simultaneously,the convolutional block attention module (CBAM)is integrated as a lightweight attention mechanism.Experimental results demonstrate that the proposed U-Net3+ model delivers significant improvements in segmentation performance,achieving an 8.3% increase in overall accuracy,2.6% in mean intersection over union,and 1.9% in F1-score compared to the original U-Net3+ model.The proposed model consistently outperforms established benchmarks,including FCN,U-Net,U-Net++,and the DeepLab series,across all evaluation metrics,demonstrating superior feature discrimination and segmentation accuracy in representative scene types.Moreover,the integration of either ResNet50 or CBAM alone results in moderate gains,their combined implementation leads to a notable synergistic effect,yielding the most effective results in segmentation tasks.The improved U-Net3+ model has significantly improved the segmentation accuracy,providing an effective technical solution for semantic segmentation of UAV orthophoto maps.

Key words: UAV orthophoto images, semantic segmentation, U-Net3+, ResNet50, convolutional block attention module