Effect of data augmentation on few-shot classification performance of surface defects in cold-rolled steel strips

Few-shot classification of surface defects in cold-rolled steel strips is a significant challenge in industrial quality inspection, with the core difficulties being sample scarcity and class imbalance. To systematically investigate the efficacy of different solution strategies, this paper constructs a large-scale dataset comprising 53 categories and 14 499 samples, and systematically compares the impact of traditional transformation-based augmentation and generative augmentation methods－such as Projected GAN and Diffusion－on the performance of classification models. Our findings reveal that although both Projected GAN and Diffusion are capable of generating high-quality images, with Diffusion demonstrating strong cross-domain migration capability, their augmentation effects are limited. Generative methods only slightly improve accuracy at the cost of a decline in recall and F1-score, while traditional image transformation-based augmentation even leads to negative effects. Surprisingly, a simple balanced sampling strategy achieves the best performance. This study confirms that directly addressing class imbalance is crucial for improving few-shot defect classification performance. Furthermore, it provides important empirical evidence for selecting technical solutions in industrial inspection scenarios: compared to complex generative augmentation, lightweight balanced sampling may represent a more efficient and practical choice.