Manual detection of underwater dam damages is time-consuming, labor-intensive, and dangerous. This paper introduces an automatic and real-time model for multiple dam damage detection based on deep learning methods. To efficiently detect the images or videos captured by remotely operated vehicles, you only look once version 8 was modified to improve the detection performance. A specialized small object detection layer, P2, was incorporated to identify minor damages. Additionally, a channel prior convolutional attention mechanism was implemented preceding each detection head to refine feature processing. The model was validated on a custom dataset demonstrating superior performance in detecting various types of underwater dam damages compared to other models, achieving mAP50 of 0.842. This approach enables efficient, intelligent, and real-time underwater dam damage detection and holds the potential to significantly assist engineers in assessing dam integrity.