【摘要】 针对城市在建道路背景的复杂性以及无人机获取RGB数据源的单一性导致的道路多分类精度较低的问题，提出了一种改进Swin-Unet模型的RGB与DSM三分支多级融合方法。通过引入Swin Transformer结构的并行编码器，增强了对RGB与DSM影像特征的提取能力；在并行编码器间加入Concat分支，提升了特征信息的融合效果；引入多密集跳跃连接，实现了不同层级特征融合，提高了模型的分割精度。实验结果表明，相较于传统的RGB与DSM影像融合方法，所提出的改进Swin-Unet模型的融合方法在总体分类精度(94.86%)、召回率(94.39%)和F1分数(94.54%)等评价指标上均表现出优越性，证明了该方法能够有效应用于在建道路各阶段信息提取，为在建道路项目的施工进度监测提供方法支持。更多还原

【Abstract】 Aiming at the problem of low classification accuracy of road multi-classification caused by the complexity of urban road under construction and the singleness of RGB data source obtained by UAV,this paper proposes a three-branch multi-level fusion method of RGB and DSM based on the improved Swin-Unet model. By introducing the parallel encoder of Swin Transformer structure, the extraction ability of RGB and DSM image features is enhanced; by adding Concat branch between parallel encoders, the fusion effect of feature information is improved; by introducing multiple dense skip connection, the fusion of different levels of feature information is realized, and the segmentation accuracy of the model is improved. Experimental results show that compared with the traditional RGB and DSM image fusion method, the fusion method of the improved Swin-Unet model proposed in this paper has superiority in overall classification accuracy(94.86%),recall rate(94.39%)and F1 score(94.54%)and other evaluation indexes, which proves that the method can be effectively applied to the information extraction of each stage of road under construction, and provides method support for the construction progress monitoring of road projects under construction.更多还原