Crack Segmentation with Generative Deep Learning-Based Data Augmentation Approach

Kwanghee Won, Mostafa Tazarv, Timothy M. Hansen

Abstract: This paper presents a novel data augmentation approach for pixel-level crack detection (segmentation) task. The proposed approach employs deep generative models in three subtasks. Firstly, a generative adversarial network (GAN) is utilized to generate crack-free images from images containing cracks and their corresponding crack masks. Secondly, another GAN is employed to learn and generate crack masks. Thirdly, a GAN is trained to generate realistic-looking cracks on a crack-free image using the crack mask obtained from the second GAN. The experimental results demonstrated that the proposed data augmentation method improved the performance on the validation set for both UNet and Attention UNet models.