New AI method generates synthetic defect data for industrial surface inspection

A recent research article presents an unsupervised method for identifying surface defects in industrial settings, integrating a Denoising Diffusion Probabilistic Model with an asymmetric teacher-student network design. This approach tackles prevalent issues such as the scarcity of defect samples, extreme long-tailed distributions, and the challenge of pinpointing subtle defects against intricate backgrounds. Initially, the DDPM is trained solely on normal samples, producing high-quality defect samples with pixel-level annotations through Gaussian perturbations and Perlin noise masks. This synthetic data generation mitigates data limitations while ensuring physical coherence. The model features an asymmetric dual-stream network, where the teacher network offers stable normal feature representations, while the student network enhances the differences between normal and defective areas. This technique seeks to enhance accuracy in industrial inspections where real defect data is limited or unbalanced. The paper was recently published on arXiv under the identifier 2604.19240v1.