Industrial defect image generation via structured latent-space interpolation

Abstract Industrial visual defect detection suffers from scarce defect samples, limiting generalization. We propose a novel image generation framework that synthesizes diverse and realistic defect images through three sequential innovations. First, we introduce UMAP-based dimensionality reduction to project VAE latent features onto a low-dimensional manifold while preserving local neighborhood structure, thereby enabling artifact-free sampling in the subsequent stage. Second, we design a PSO-optimized Gaussian mixture model to capture the sparse, multimodal distribution of defect features within this manifold, and sample new coordinates from high-likelihood regions to generate diverse yet realistic defect representations. Third, we employ a non-parametric radial basis function interpolator to map these sampled coordinates back to the original VAE latent space, preserving geometric continuity and restoring high-frequency defect details before decoding. Experiments on bearing-defect and public industrial datasets show that detectors trained with a synthetic-to-real ratio $$\le 1:3$$ achieve superior accuracy and robustness, validating each innovation’s effectiveness.