The complex apparent environment of concrete dams causes crack detection difficult. To address the problems of low precision of crack detection by traditional methods, this paper proposes a concrete dam apparent crack detection algorithm based on the improved DeepLabv3+ model, which provides more accurate and intuitive detection results for concrete dam health monitoring. First, a self-made concrete dam crack dataset is constructed for model training and testing. Then, the optimal initial learning rate is selected by comparing and analyzing in five different learning rates. Finally, on the test set, the model proposed in this paper performs well, with high IOU and PA. It can conclude that the proposed method efficiently and accurately detects the concrete dam cracks in complex backgrounds, providing effective technical support for the long-term monitoring of concrete dams.