The research presents a valuable contribution to the detection of surface defects of fair-faced concrete through proposing an image augmentation method with a Wasserstein generative adversarial network. Integrating a self-attention mechanism and gradient penalty, the algorithm substantially mitigates data scarcity by generating high-quality and diverse defect images. Trained by data collected from five structures, the model improves recognition accuracy, which can serve as an effective approach for small sample learning in the analysis of concrete defects.