Abstract:

[Objective] In underwater engineering inspection, the turbid shallow water environment severely hinders the performance of machine vision-based methods for detecting surface defects in underwater structures. To address the challenge of defect detection in turbid water, this study proposes a lightweight three-stage underwater defect detection method that integrates polarization imaging and deep learning techniques. A defect detection model, named PCC-YOLOv7, is developed. [Methods] First, polarization imaging technology was combined with a polarization restoration model to analyze the polarization characteristics of light waves. This approach effectively suppressed scattering interference in turbid water, thereby achieving clear imaging of turbid environments and restoring defect images. Consequently, defect details obscured by scattering particles were reconstructed. Second, the CAA-SRGAN (Coordinate Attention ACON-Super Resolution Generative Adversarial Network) model was introduced. By employing an improved attention mechanism and a generative adversarial network structure, super-resolution processing was performed on the restored images. This yielded high-resolution underwater defect images, providing a high-quality data foundation for subsequent precise detection. Finally, a defect detection model based on CBAM-YOLOv7 was established, where the convolutional block attention module (CBAM) was utilized to enhance the network’s focus on defect features. Leveraging the advanced YOLOv7 object detection framework, common underwater structural defects, including cracks, holes, and spalling can be rapidly and accurately identified. These three sub-models worked collaboratively to form a comprehensive detection system. [Results] For image restoration, the polarization restoration model exhibited superior performance in metrics such as image clarity and color fidelity compared to current restoration methods. The CAA-SRGAN model generated images with notable improvements in detail texture preservation and resolution enhancement. The CBAM-YOLOv7 defect detection model achieved higher accuracy in both defect localization and classification. A comprehensive evaluation of the PCC-YOLOv7 defect detection model revealed an average improvement of 33.5% in mean average precision (mAP_0.5, mAP_0.75, and mAP_0.5-0.95). Compared to existing models, PCC-YOLOv7 significantly enhanced defect detection performance in turbid underwater environments, effectively improving both recognition rate and detection efficiency. [Conclusions] The PCC-YOLOv7 defect detection model innovatively integrates polarization imaging technology with deep learning. Through the collaborative operation of three functionally complementary sub-models, it successfully addresses the challenge of detecting surface defects in underwater structures in turbid water. Compared to existing models, the proposed model demonstrates enhanced adaptability to turbid underwater detection scenarios. It enables stable and efficient detection of surface defects in underwater structures under complex turbid conditions, providing a practical technical solution for the safety assessment and maintenance of underwater structures. Future work may focus on further optimizing the model structure and extending its application to more underwater scenarios.

Key words: turbid water, underwater structure, defect detection, polarization imaging, deep learning, super-resolution reconstruction