A robust vision-based method for damage detection of underwater structures using image enhancement technique

Abstract

Underwater structures are highly susceptible to damage from various factors, such as corrosion, erosion, and scouring, which can significantly reduce their load-bearing capacity. Detecting structural damage in these environments is crucial. Computer vision technology has been widely applied to the damage detection of infrastructures because of its simple operation and noncontact advantages. The captured images of underwater structures are often blurred due to the complexity of the underwater environment, significantly affecting damage detection. This study proposes a robust vision-based method for damage detection in underwater structures. Specifically, considering the potential uncertainty, the PUIE-Net model is introduced to generate multiple enhanced images and construct the corresponding probability distribution. The generated images with the highest probability distribution are selected as the optimal enhanced images, ensuring that the enhanced images are similar to the real waterless images. Subsequently, an attention mechanism is introduced to improve the YOLOv11 model. Additionally, dataset augmentation techniques are employed to simulate various underwater environments to expand the datasets for model training. Field experiments are conducted on an actual concrete bridge to validate the effectiveness of the proposed method. The damage images of the underwater structures enhanced by the proposed method are clearer than those enhanced by other methods. The proposed improved YOLOv11 model detects structural damage in enhanced underwater images with higher accuracy than the original model. In addition, the proposed method demonstrates strong robustness in detecting damage in underwater structures, even under varying camera shooting angles, small-scale damage, and camera shaking.