Transfer Learning and Super Resolution Assisted Few-Shot and Small Defect Detection in Printed Circuit Boards

Defects on printed circuit boards (PCBs) can cause circuit connection failures, significantly compromising the reliability of electronic devices. Traditional defect detection methods require extensive annotated data, limiting their practical applications in few-shot and small-object scenarios. Therefore, this paper proposes a transfer learning framework for few-shot and small-object PCB defect detection. First, we introduce a defect fusion enhancement strategy and a cross-domain defect generation method to expand the few-shot dataset. The former controls the location of defect templates, while the latter uses CycleGAN and public PCB defect datasets to augment existing defect templates. Second, we design a super-resolution-assisted transfer learning strategy, which incorporates a super-resolution reconstruction branch during the training phase to achieve efficient knowledge transfer. Experimental results show that the cross-domain defect generation method successfully produces a high-quality synthetic dataset. Fine-tuning on this synthetic dataset significantly improves detection performance in the tested few-shot PCB scenarios, while the integration of the super-resolution branch improves the model’s perception of small objects.