【摘要】 为提高建筑损害的评估精度，提出一种基于UCTransNet的双阶段灾后建筑损害评估模型(MGDLNet)。阶段一使用UCTransNet完成建筑分割。阶段二使用改进后的DM-UCTransNet进行建筑损害评估，通过差异特征提取模块充分融合多尺度的建筑损害特征，嵌入空间金字塔更好捕捉小目标建筑及边缘特征，引入深度监督机制和改进损失函数加强浅层特征学习并平衡样本。实验结果表明，MGDLNet在目标数据集有较大优势，其加权F1得分相较于SegNet、UNet、DeeplabV3+、TransUNet和UCTransNet分别提高了8.6%、1.9%、5.0%、2.7%和1.4%。更多还原

【Abstract】 To improve the accuracy of building damage assessment, a two-stage post-disaster building damage assessment model(MGDLNet) based on UCTransNet was proposed. In the first phase, UCTransNet was used to complete the building subdivision. In the second stage, the improved DM-UCTransNet was used for building damage assessment. The differential feature extraction module was used to fully integrate multi-scale building damage features, and the spatial pyramid was embedded to better capture small target buildings and edge features. The deep supervision mechanism and improved loss function were introduced to strengthen shallow feature learning and balance samples. The results of experiment show that MGDLNet has a great advantage in the target dataset, and its weighted F1 score is increased by 8.6%, 1.9%, 5.0%, 2.7% and 1.4% compared with that of SegNet, UNet, DeeplabV3+, TransUNet and UCTransNet, respectively.更多还原