Crack Structure Preserving Environmental Appearance Augmentation for Robust Surface Crack Segmentation in Complex Environments

Crack segmentation is an important task in infrastructure maintenance. Engineering structures require regular crack inspection and maintenance to reduce the risk of further deterioration. For crack segmentation under different environmental conditions, sufficient and representative training data are needed to train neural network models effectively. However, variations in illumination, visibility, surface texture, image resolution, and environmental noise during practical image acquisition can alter crack appearance and reduce segmentation stability. To address these problems, this study develops a sequential crack data augmentation framework using GAN, CNN, and VGG19. The GAN first generates diverse crack images. A mask prediction CNN then produces the corresponding binary pseudo-label masks, after which VGG19 style transfer introduces the target environmental appearance. During style transfer, the content constraint limits changes to crack-related structure, while the previously predicted mask remains unchanged and is paired with the transferred image. The resulting image and mask pairs are used to expand the training data for subsequent crack segmentation. The framework is evaluated using crack datasets representing different environmental conditions. Experimental results show that the proposed augmentation procedure supports crack segmentation under complex environmental interference and improves the model’s adaptability to variations in environmental appearance. The framework provides a practical approach for expanding crack training data when representative images from complex environments are limited.