Enhancing automotive surface defect segmentation with physics-inspired synthetic data

Abstract Semantic segmentation of metallic surface defects plays a critical role in industrial visual inspection but is severely constrained by the scarcity and high cost of pixel-level annotated real defect data. Synthetic data generation offers a promising alternative; however, its practical contribution to defect segmentation under realistic evaluation settings remains insufficiently understood. This study systematically investigates the role of synthetic data in metallic surface defect segmentation, with a focus on data efficiency under limited real defect availability. A geometry-based, physics-inspired synthetic data generation pipeline is employed to produce localized metallic surface images containing parameterized bumps, dents, and scratches under diverse illumination conditions, providing pixel-accurate ground-truth masks by construction. Synthetic images are combined with real defect-free and defective images to form multiple training regimes with increasing numbers of real defective samples. A U-Net-based semantic segmentation model is trained on each regime and evaluated using a fixed real-only test set to ensure unbiased comparison. Performance is assessed using both pixel-wise and defect-wise evaluation criteria. Experimental results show that synthetic augmentation substantially improves segmentation accuracy and defect detection completeness in low-data regimes, yielding higher overlap-based metrics and more reliable instance-level detection when real defective data are scarce. As the amount of real defective data increases, the relative contribution of synthetic data diminishes, though synthetic augmentation consistently maintains higher Dice scores. These findings indicate that synthetic data serve as an effective complementary training source for industrial defect segmentation, particularly in data-limited scenarios, rather than a replacement for real-world samples.