Abstract: The detection and identification of surface cracks on dams is of great significance for dam safety, so we study dam surface cracks detection based on deep learning.In view of the complex topological structures and imbalance of positive and negative samples of the crack images, the ASPP and CBAM optimization modules were embedded in the typical U-net model, and a Dice+BCE hybrid loss function was used to replace the single cross entropy loss function.The improved U-net model performed well on a self-made instance dam crack image dataset, with IoU being 47.05% and F1 being 62.99% respectively.Compared with typical U-net model, it had increased by 5.41% and 5.19%,and compared with PSPNet model, it had increased by 3.05% and 3.31% respectively.The improved U-net model provides more accurate pixel classification and richer multi-scale information in crack segmentation tasks, providing a better means for detecting and identifying surface cracks in dam concrete structures.