DAFF-Net: A Detail-Aware Downsampling and Enhanced Feature Fusion Network for Underwater Dam Crack Image Object Detection

With the increasing demand for dam safety monitoring, underwater dam crack detection has become a critical task for ensuring structural integrity. The rapid development of computer vision technologies offers efficient, intuitive, and noncontact solutions for this task. However, the complex underwater environment presents significant challenges, making traditional detection methods inadequate in both accuracy and efficiency for practical applications. In response, we propose DAFF-Net, a novel model tailored for underwater dam crack detection. We propose a Detail-Aware Downsampling Module that effectively preserves essential fine-grained texture information while reducing spatial resolution to alleviate information loss. Furthermore, we design a Serial-Cascaded Feature Fusion Module to fully integrate multiscale features, enhancing the model's capability to capture both local details and global contextual information. In addition, we introduce an optimized loss function SlideLoss to guide the model in better capturing the spatial structure of crack regions, thereby enhancing model performance. Experimental results demonstrate that the proposed DAFF-Net achieves excellent performance on underwater crack detection datasets. This method achieves both high accuracy and efficient inference speed, providing reliable technical support for intelligent dam safety monitoring in practical engineering applications.