针对修建在高原恶劣环境下的的隧道洞门墙病害频发及传统人工病害检测方法效率低下、危险性高等问题，提出了基于无人机图像与深度学习的高原区隧道洞门墙病害检测方法。首先使用无人机航拍采集了新疆某高原区隧道洞门墙病害图像，建立一个多病害语义分割数据集。随后以语义分割模型DeeplabV3+为基础，提出以MobileNetV2作为主干特征提取网络的改进模型TP-DeeplabV3+，以减少模型参数；采用FocalLoss作为损失函数，以解决病害图像中的类别不平衡问题；添加CA注意力机制，以提升模型分割性能；最后提出病害量化方法。实验结果表明，TP- DeeplabV3+在测试集上可以达到88.37%和94.93%的mIoU和mPA，模型体量压缩了88.83%；提出的病害量化方法对于病害覆盖率的绝对误差不超过0.3%，相对误差维持在7.31%以下。相比传统方法，所提方法安全、准确地实现了高原恶劣环境下的隧道洞门墙智能化检测。

Aiming at the frequent diseases of tunnel portal built in the harsh environment of plateau region and the problems of low efficiency and high risk of traditional manual disease detection method, a disease detection method of tunnel portal in plateau region based on UAV image and deep learning was proposed. Firstly, an Unmanned Aerial Vehicle (UAV) was used to collect the disease images of a tunnel portal in the plateau region of Xinjiang, and a multi disease semantic segmentation dataset was constructed. Then, based on DeeplabV3+, an improved model TP-DeeplabV3+ was proposed, which used MobileNetV2 as the backbone feature extraction network to reduce the model parameters; Focalloss was used as the loss function to solve the problem of category imbalance in disease images; CA attention mechanism was uesd to improve the segmentation performance; and the disease quantification method was proposed. Experiment results show that TP-DeeplabV3+ can reach 88.37% and 94.93% of mIoU and mPA on the testset, and the model volume is reduced by 88.83%; The absolute error of the proposed disease quantification method for disease coverage rate is less than 0.3%, and the relative error is maintained below 7.31%. Compared with the traditional manual method, the proposed method can realize the intelligent detection of tunnel portal safely and accurately in plateau region.

新疆重大科技专项(2020A03003-7）；陕西省自然科学基础研究计划面上项目(2021JM-180)；中央高校基本科研业务费资助项目 (领军人才计划)(300102211302)