Abstract
Few-shot steel surface defect generation remains challenging due to the limited availability of training samples and the complex visual characteristics of industrial defects. Traditional data augmentation techniques often fail to capture the diverse manifestations of defects, resulting in suboptimal detection performance. While existing stable diffusion models possess robust generative capabilities, they encounter domain knowledge transfer limitations when applied to industrial settings. To address these constraints, stable industrial defect generation (stableIDG) is proposed, which enhances stable diffusion through three key improvements: (1) fine-tuning with low-rank embedding adaptation to accommodate steel surface defect generation; (2) implementation of personalized identifiers associated with defect regions to prevent malformation during generation; and (3) development of a mask-guided defect learning mechanism that directs network attention toward critical defect regions, thereby enhancing the fidelity of generated details. The method was validated on a newly constructed Medium and Heavy Plate Surface Defect Dataset (MHPSD) from a real industrial environment. Experiments demonstrated that stableIDG achieved better image generation quality compared to existing methods, such as textual inversion. Furthermore, when the generated samples were used for data augmentation, the detection network performance was effectively enhanced. Detection recall significantly improved for both defect classes, increasing from 0.656 to 0.908 for Inclusion defects and from 0.86 to 0.986 for Foreign Object Embedding defects. These results demonstrate the effectiveness of the proposed method for industrial defect detection in data-scarce scenarios.