Abstract: The aim of this study is to improve the quality and quantity of the dataset for surface defect images of underwater structures and facilitate the application of deep learning methods in underwater detection. A method for generating surface defect images of underwater structures is proposed based on the deep convolutional generative adversarial networks (DCGAN). First, the image quality is guaranteed by designing an underwater image acquisition device through the adjustment of shooting distance and the supplement of light intensity. Second, by improving the loss function and optimizing DCGAN, the image generation model for surface defect of underwater structures is established. Finally, the effectiveness of the generated images is assessed using the YOLOv5 detection network. The results demonstrate an average peak signal-to-noise ratio of 21.142 6 dB and an average structural similarity of 0.716 8 for the generated crack images of underwater structures. Integrating the generated and real images into the detection model effectively improves the accuracy of detection. The study provides technical support for the health detection of hydraulic structures such as dams and headrace tunnels.

Key words: underwater structure, surface defect detection, deep learning, image generation, deep convolutional generative adversarial networks