A real-time crack detection approach for underwater concrete structures using sonar and deep learning

This paper introduces a real-time crack detection approach for underwater concrete structures using sonar and deep learning to overcome limitations in low-light or turbid environments where optical imaging struggles. Specifically, a crack detection model based on the YOLOv5s architecture was developed for sonar images, incorporating attention mechanisms and the SIoU loss function to improve detection accuracy. Given the scarcity of acoustic crack image data, a two-stage transfer learning approach was implemented, leveraging both source domain data (publicly available optical crack images) and target domain data acquired from on-site acoustic detection experiments. Ablation studies and comparisons with other advanced models indicate that the proposed model achieves robust detection accuracy (mAP@0.5 = 0.768) with an inference speed of 134 FPS, making it suitable for real-time applications. Additionally, a pixel-based analysis method was used to estimate overall crack dimensions, providing valuable insights into crack characteristics and their potential structural impact.