Robust PCB Anomaly Detection Using Aligned Design Information

Abstract

Printed circuit board (PCB) is a critical component in modern electronic products, and ensuring its high quality is essential for optimal device functioning. Unsupervised learning-based anomaly detection methods can identify PCB defects using only defect-free samples, reducing labor costs. However, these methods may overlook valuable reference information from Gerber—the PCB design information. Leveraging this information has the potential to greatly enhance the precision of identifying PCB defects. We propose the first PCB anomaly detection framework utilizing Gerber. Specifically, we use a fixed real image extractor to extract features from real images, and then transform the extracted Gerber features into real image features through a Gerber feature extractor and transformer. By comparing the transformed features with the real image features, we achieve effective PCB anomaly detection. Furthermore, taking into account potential discrepancies in offset between Gerber and real images, as well as the presence of noise in data collection, we have introduced an offset-tolerant matching (OTM) algorithm and a noise-resilient scheme to bolster the robustness of the model. Through experiments conducted on industrial data collected from real production environments, our proposed approach achieves a performance of 90.46% AP and 96.23% AUROC, demonstrating state-of-the-art results.