High-resolution concrete damage image synthesis using conditional generative adversarial network

Abstract

Concrete damage images are essential for training deep learning-based damage detection networks. Considering the manual collection of concrete damage images is time-consuming and labor-intensive, this study proposes a synthesis method for high-resolution concrete damage images using a conditional generative adversarial network (CGAN). To this end, pix2pix, CycleGAN, OASIS, and pix2pixHD with various hyperparameters were trained and tested on 500 concrete crack and spalling images. The test results show that the trained pix2pixHD with λpix2pixHD = 15 is the best CGAN for concrete damage image synthesis. Concrete damage images were synthesized by the best CGAN according to hand-painted damage maps and used to train deep learning networks. The results show that the synthesized images have excellent authenticity and can be used to train and test deep learning-based concrete damage detection networks. The proposed method can be enhanced by adding damage images to the existing database or employing a better CGAN generator.