Aiming at the problems of harsh environment, frequent frost damage, multiple interference factors in lining images, inconsistent scale of frost damage targets, and low efficiency and high cost of traditional manual visual inspection methods for tunnel lining built in cold regions, a frost damage detection method for tunnel lining in cold regions based on HRNetV2 was proposed. Firstly, based on HRNetV2, an improved model was proposed, which combines the knowledge of transfer learning in the backbone feature extraction network, introduces attention mechanism in the structure to enhance the model's learning ability for frost damage features, and uses Focalloss as the loss function to solve the problem of class imbalance. In order to verify the performance of the improved model, frozen damage images of tunnel lining in cold regions were collected by high-definition cameras. After clipping and data enhancement, a semantic segmentation dataset containing 2 800 cracks, spalling and ice hanging was established. The experimental results show that the mean intersection over union(mIoU) of the improved model can reach 89.05% on the freezing damage dataset, which is 5.41% higher than that of the original model, showing good robustness in the face of complex forms of frost damage, and can be directly applied to high-resolution original images. Compared with DeeplabV3 + , U-Net and PSPNet three classical semantic segmentation models, the improved HRNetV2 model has more advantages in comprehensive performance. The proposed method can accurately and safely achieve intelligent detection of lining frost damage, providing certain technical support for intelligent operation and maintenance of tunnels in cold regions.