A multi-stage method for defect detection of underwater structures based on deep learning

Underwater defect detection faces challenges such as difficulty in image acquisition, low precision in detection and inaccurate defect localization. This article presents a multi-stage method to address these issues. A custom-built remotely operated vehicle (ROV) with advanced path planning was used to collect images of underwater defects. An improved YOLOv8 network, integrating deformable convolution and a multi-head self-attention mechanism, significantly enhanced defect detection accuracy. Furthermore, an upgraded Deeplabv3+ semantic segmentation network with a densely connected atrous spatial pyramid pooling module was proposed for precise pixel-level mapping of defects, particularly elongated ones. A 3D reconstruction method based on structure from motion was developed to generate accurate 3D point clouds for precise defect localization. The experimental results demonstrated that the developed ROV, equipped with a high-resolution camera and a multi-source heterogeneous vision enhancement module, efficiently captured defect images and improved image quality in turbid water. The improved YOLOv8 achieved a 6.61% increase in mAP50, while the upgraded Deeplabv3+ showed a 4.19% increase in mean intersections over union. These enhancements enabled the integrated method to achieve pixel-level defect detection and segmentation, demonstrating significant advancements across all performance metrics and competitive frames per second for real-time applications. The successful visualization of defects in the 3D model validated the effectiveness and feasibility of the proposed multi-stage method.