Dynamic patch bank with a defect filter for cold start in online engineering inspection

Abstract

In industrial defect detection, real-world complexities such as cold starts, pervasive label noise, and newly emerging defect classes pose significant challenges to traditional approaches. We propose a Dynamic Patch Bank with a Defect Filter, a unified artificial intelligence based framework that addresses these issues simultaneously. First, a local embedding density–based filter isolates truly defective patches from globally labeled samples, mitigating noise and enabling cold-start learning. Next, a dynamic patch bank continuously refines representative embeddings, while an elastic weighted feedback mechanism integrates novel defects without retraining or catastrophic forgetting. Our approach introduces a novel machine learning based feature extraction and anomaly detection mechanism to enhance defect classification. On the MVTec Anomaly Detection (MVTec AD) dataset, our method uses fewer than four samples per defect type to achieve leading performance. Further evaluations on four additional benchmarks confirm its robustness. Code is publicly available at https://github.com/YoungerAnonymous/DPB-DF .