Underwater dam crack image generation based on unsupervised image-to-image translation

Abstract

Underwater crack detection is necessary for the safe operation of concrete dams. Current deep-learning-based underwater crack detection methods rely heavily on a large number of crack images; these images are difficult to collect because of complex and dangerous underwater environments. This study proposes a method that can generate underwater dam crack images using above-water dam crack images through an image-to-image translation method. CycleGAN, which is more powerful than other models, was optimized to generate underwater dam crack images by translating existing above-water dam crack images into underwater styles. The generated underwater images were validated using the following major computer vision tasks: classification, object detection, and semantic segmentation. The results demonstrate the potential of these generated images for training and optimizing deep-learning-based crack-detection methods. This method is expected to solve the problem of insufficient data and enhance deep-learning-based detection networks.