Few-shot GAN adaptation for high-fidelity and diverse crack image generation in dam damage detection

Abstract

Substantial crack imagery is hard to acquire in dam structural inspection due to high costs and risks. Crack image generation, as a crucial yet challenging visual task, still struggles with the quality-diversity trade-off under data scarcity. This paper thus presents CrackFSGAN, a few-shot Generative Adversarial Network (GAN) adaptation method for generating realistic, diverse dam crack images from limited samples. It incorporates the Cross-Scale Channel Interaction (CSCI) module to ensure robust gradient flow across network weights for efficient training, and the Self-Supervised Discriminator (SSDr), a redesigned feature-encoder with an additional decoder, to learn more discriminative, region-extensive feature maps. Extensive experiments on multiple damage datasets against state-of-the-art GANs validate CrackFSGAN's superiority in few-shot image synthesis quality and diversity, and its effectiveness in data augmentation for downstream crack detection tasks. Notably, it supports high-resolution (1024 × 1024 pixel²) crack image generation, offering a promising solution to data scarcity and advancing intelligent structural damage detection.