To tackle the difficulty of detecting underwater cracks in concrete dams effectively with deep learning algorithms, a semantic segmentation model for underwater cracks in concrete dams based on MobileNetV2-DeepLabv3+ was proposed. The lightweight network MobileNetV2 was introduced into the model and the deep feature downsampling multiplier was reduced to 8, so as to improve the recognition accuracy and inference speed under small dataset conditions. To alleviate the problem of category imbalance, the combination of cross entropy loss function and Dice loss function was used as the loss function of the model. The validation results from an engineering case show that the mean pixel accuracy and mean intersection over union of the model on the test set are as high as 90.87% and 86.33%, respectively, meeting the requirements for underwater cracks of high-precision semantic segmentation. Compared with other models, the proposed model shows better segmentation effect of underwater cracks of concrete dams under typical conditions, with strong generalization ability. With features of low memory usage and high inference speed, the proposed model is suitable for underwater crack detection of concrete dams.