基于对抗式域适应迁移模型的震害建筑物损毁识别方法研究

准确获取震害建筑物损毁信息对指导灾后救援具有极其重要的意义。作为迁移学习的分支，域适应理论在利用历史场景震害影像解决当前损毁识别中缺少标记样本的问题上展现出潜力。然而，震害损毁建筑物复杂多变，不同场景图像的差异较大，同一场景图像的类间相似度较高，类内差异性较大，严重影响了域适应的迁移性能，进而影响了识别准确率。为此，本项目拟提出对抗式域适应迁移模型，从域间整体特征适配、类别特征适配、困难样本处理三个层面建立鲁棒的迁移机制。首先研究基于生成对抗网络的对抗式域适应模型的构造与高效学习方法，实现不同场景图像的整体特征适配；然后提出变分贝叶斯类别特征适配机制，改善域适应模型的负迁移现象；接着研究基于截断高斯混合模型和类别概率估计的困难样本处理方法，进一步提升模型的特征适配效率；最后基于所提出的迁移模型，研究震害建筑物损毁的识别方法。研究成果有望突破建筑物损毁评估的技术瓶颈，推动迁移学习理论的发展。

Accurate assessment of building damage is critical to rescue operations in the aftermath of earthquake. As one of the branches of transfer learning, leveraging the images of historical earthquake-induced building damages, domain adaptation exhibits great potentials in solving the issue of insufficient labeled samples in the current damage identification task. However, the images of damaged buildings are complicated, which have characteristics of large discrepancy between different scene images. Even for the images of the same scene, they exhibit high similarity of inter-class and large diversity of intra-class. All these factors greatly decrease the transfer performance of domain adaptation, as well as the recognition accuracy. To address these issues, in this proposal we will build an adversarial domain adaptation transfer model, and develop a robust transfer mechanism from the view of cross-domain feature adaptation, category-specific feature adaptation and processing of hard samples. First, domain adaptation model based on Generative Adversarial Networks (GAN) will be constructed, and the efficient learning strategy will be studied which can realize feature adaptation of cross-domain. Then, a Variational Bayes based category-specific feature adaptation mechanism will be proposed to alleviate the effect of negative transfer. And, methods of hard sample processing based on truncated Gaussian Mixture Model and category probability estimation will be developed, which can further improve the efficiency of feature adaptation for the transfer model. At last, the building damage recognition method based on the proposed transfer model will be investigated. The research results will be potential to break the bottle neck of building damage assessment, and will be beneficial to the development of transfer learning theory.