FlawMatch: Conditional defect image generation via flow matching for improved surface defect classification

Abstract

Developing robust binary classification models for surface defect detection is essential for ensuring quality and reliability in industrial inspection systems. A persistent challenge in this context is data imbalance, wherein positive (e.g., defective instances) samples are scarce and underrepresented. This scarcity further limits the diversity of rare defect patterns, thereby impeding the generalization capabilities of classifiers. Although generative models have been employed to augment minority-class data, many existing approaches fail to synthesize realistic and diverse defect patterns, limiting their effectiveness for classifier training. This study presents FlawMatch, a class-aware conditional generation framework that synthesizes realistic defect images using flow matching. FlawMatch captures spatial and geometric attributes from real defects and encodes them into discrete class labels, which subsequently condition a flow matching model. The generated defects are composited onto defect-free backgrounds to construct full training images. Experiments on both proprietary semiconductor and public (KolektorSDD2) datasets demonstrate the superiority of FlawMatch over state-of-the-art diffusion-based methods. Specifically, FlawMatch demonstrates a 37% improvements in both Fréchet Inception Distance and 43% in Kernel Inception Distance, indicating a more accurate generation of image distributions. Furthermore, it improves the downstream classification performance, yielding average gains of 2% in recall and 1% in F1 score, while also achieving a 17× reduction in sampling time compared to DDPM-based diffusion models. In summary, FlawMatch delivers an efficient and structurally grounded solution for defect synthesis under conditions of data scarcity, offering practical value for industrial surface inspection applications.