Pseudo defect guided meta calibration for few shot industrial defect segmentation

Abstract

Few-shot industrial defect segmentation aims to localize unseen defects from only a few defect-free reference images, but remains challenging because no positive anomaly examples are available and defects are often subtle, irregular, and embedded in complex textures. We propose PDMCNet, a Pseudo-Defect Guided Meta-Calibration Network for normal-reference few-shot industrial defect segmentation. PDMCNet synthesizes task-specific pseudo defects from normal support images and uses them as surrogate anomaly cues to calibrate the model at inference. Rather than adapting the whole network, it updates only a scalar meta-bias and a set of lightweight projection layers in the inner loop to adjust foreground activation near the decision boundary while preserving the learned representation. A dense correlation and reconstruction module further reconstructs query features from normal support features at the pixel level, enabling residual-based localization of subtle defects. Experiments on the Industrial-5 \(^i\) benchmark show that PDMCNet achieves mIoU scores of 36.03% and 38.98% under 1-shot and 5-shot settings, respectively, outperforming the compared few-shot segmentation and anomaly segmentation baselines. Ablation and sensitivity analyses demonstrate that pseudo-defect guided scalar calibration provides a lightweight, stable, and interpretable mechanism for adapting to unseen industrial categories.