SIOWGAN: a single-image-based diverse virtual sample generation method for surface defect detection

Surface defect detection plays a crucial role in industrial manufacturing. Due to the scarcity of defect samples, traditional deep learning models face significant limitations in detection performance and generalization ability. To address this issue, this paper proposes a single-image orthogonal weights generative adversarial network (SIOWGAN), a multi-generator adversarial generative model designed to effectively augment defect samples under limited sample conditions. This method requires only a single sample for data generation. The multi-generator structure, together with weight-orthogonality regularization, mitigates mode collapse and improves sample diversity under low-data conditions. Compared with existing multi-generator architectures, we propose a novel weight orthogonality training strategy, which guides different generators to learn distinct data distribution modalities without introducing any additional models. This design avoids the computational overhead and training instability that typically arise from incorporating extra models. Experimental results on the NEU-DET metal surface defect dataset demonstrate that the proposed method can generate diverse defect samples that align with the real data distribution using only 4% of the original data. The Fréchet inception distance and IS scores of the proposed method outperform those of existing multi-generator and single-image generation baseline models. The augmented samples are used to train YOLOv10, YOLOv11, YOLOv12, and YOLO26 detectors. Detection performance improves substantially, and in several settings mAP@0.5-0.95 approaches or exceeds the model trained on the full training split. Additional ablation studies validate the effectiveness of the proposed two loss functions in promoting sample diversity and maintaining realism. In conclusion, SIOWGAN provides a scalable data augmentation approach for industrial defect detection tasks under extremely limited sample conditions.