Current dam surface inspections in actual engineering applications mainly rely on manual on-site inspections or scaffolding observations to obtain the safety status of the dam surface structure. These methods have problems such as high safety risks, high costs, and low efficiency. In addition, false detection, missed detection, and subjective issues due to human eyes identification exist. The tethered UAV equipped with high-definition cameras is used to collect dam face images, which can reduce safety risks and improve efficiency. Convolutional neural networks are applied to achieve dam face image defect recognition and the accuracy of dam surface defect recognition is improved.The ResNet-152 is used as the backbone network to build a network model to extract the characteristics of cracks and defects, based on which a new decoding network layer to achieve crack pixel segmentation detection is designed. The testing results show that the crack defectsdetection precision P, recall rate R, comprehensive index F and average intersection ratio M of crack defects reach 74.61%, 78.71%, 74.99% and 73.34%, respectively. Comparative experiments were carried out with a variety of pixel-level segmentation methods, indicating that the proposed detection model can effectively identify crack defects on dam surface and provide auxiliary data support for the safety assessment of dam surface structure.