A Domain Incremental Learning Framework for PCB Continuous Defect Detection

As a critical part of most electronic equipment, defect detection of the printed circuit board (PCB) is the main concern of quality and reliability. However, different pipelines and patches of products may increase the number of defect samples with varying distributions. How to realize the inspection for distribution-shift PCB defects is of great significance for PCB continuous defect detection. To address this critical problem, we propose a domain incremental learning framework for PCB continuous defect detection (DILF-CDD), which can adapt to the new defect distribution while avoiding the catastrophic forgetting of previous defect distribution. The proposed framework captures domain-invariant defect features across domains by using domain-invariant parameters, whereas represents domain-specific features by using specific parameters. Our new optimization strategy learns domain-sharing parameters by adding a distillation loss on both the feature map and proposal levels. We also propose a novel distribution-preserving (DP) sampling mechanism that preserves the defect distribution and maximizes the number of instances to condense prior knowledge on the defect domain. Extensive experimental comparisons and ablation studies across various domain-shift settings highlight the superiority and effectiveness of our method in PCB continuous defect detection.