DINO for PCB defect detection: adaptive feature pyramid and hard negative contrastive alignment

Abstract

Detecting minute defects in printed circuit boards (PCBs) against highly repetitive background textures remains a fundamental challenge for end-to-end detectors. Subtle defect features are easily weakened during conventional crossscale fusion, while normal patterns such as traces, pads, and vias may produce strong false activations. To address these issues, we propose PCBDETR, a dual-component framework built on DINO. DINO is adopted as the base detector because its denoising training, set prediction, iterative query refinement, and explicit encoder memory provide a stable foundation for high-recall small-defect detection and memory-level feature regularization. First, an Adaptive Knowledge Feature Pyramid Network (AKFPN) performs adaptive residual enhancement with channel-wise recalibration, replacing fixed post-fusion convolutions to amplify subtle defect responses. Second, a Hard Negative Contrastive Alignment (HNCA) loss decouples high-response background features in the encoder memory through contrastive regularization, thereby suppressing pseudo-defects. On the DeepPCB dataset, PCBDETR achieves 95.3% AP50 and 49.5% mAP with 45.5M parameters and 186G FLOPs, representing a 3.9 percentage-point mAP improvement over the DINO baseline. The proposed framework provides a practical balance between accuracy and model complexity for high-precision PCB inspection.