Underwater dam crack image enhancement and crack detection based on improved diffusion model and SDI-ASF-YOLO11

Abstract

Underwater cracks are a prevalent and hazardous structural issue during dam operations. The early identification and mitigation of cracks are enabled by the effective capture and analysis of underwater structural surface images. However, complex underwater environments and imaging mechanisms often degrade crack image quality, causing blurring, color distortion, and low contrast. These factors significantly hinder the performance of existing visual detection methods. To address these challenges, this study proposed a method for underwater crack image enhancement and detection in dams by integrating an improved diffusion model and SDI-ASF-YOLO11 architecture. First, a generative diffusion model for underwater crack image enhancement (Underwater Diffusion Model, UWDM) was developed by incorporating prior knowledge of marine underwater images through transfer learning, enabling cross-domain enhancement of crack images in dam environments. An underwater imaging platform was also constructed to generate a benchmark dataset of underwater cracks. The effectiveness of the UWDM method was validated through both qualitative and quantitative analyses. Subsequently, an improved crack detection model, SDI-ASF-YOLO11, featuring with an optimized feature fusion network, was proposed to enhance the detection accuracy and segmentation quality. Case studies showed that UWDM substantially improved the visual quality of degraded crack images. The performance metrics improved by 12.41 % in IE, 9.91 % in UCIQE, and 218.45 % in UIQM, compared to the original images. The combined utilization of UWDM and SDI-ASF-YOLO11 further boosted the detection accuracy by 51.6 % and segmentation accuracy by 45.9 %. This integration of image enhancement and model optimization enables accurate dam crack detection. The proposed approach demonstrated effectiveness in crack detection, applicability in engineering environments, and superiority in terms of methodological design.