Common PCB (Printed Circuit Board) defects include missing holes, shorts, spurs, etc., which may lead to product performance degradation, malfunction or safety hazards. Within the framework of Smart Manufacturing and Industry 4.0, industry strives to achieve automated and intelligent PCB defect inspection by using advanced machine vision systems and artificial intelligence algorithms. However, PCB defect detection faces challenges such as high density and miniaturization, complex background interference, and multiscale targets. For this reason, this paper proposes a new method for PCB defect detection according to a hierarchical scale-aware attention (HSA) mechanism based on RT-DETR (Real-Time Detection Transformer), and thus the method is coded as HSA-RTDETR. The core of the new method resides in the enhancement of feature information of small target defects in a feature fusion network. Firstly, a new backbone network, R18-Faster-EMA, is designed to make the overall model more efficient; Secondly, the AIFI (Attention-based Intra-scale Feature Interaction) module is redesigned to replace the original multihead self-attention mechanism with cascaded group attention to highlight important features. Thirdly, a hierarchical scale-aware pyramid attention network (HS-PAN) is designed to realize multi-scale feature fusion and learn more comprehensive feature arrays. Finally, to improve the efficiency of the model, a new loss function is designed to speed up convergence and prioritize small target defects. Experiments show that the HSA-RTDETR method achieves a mean average precision of 96.9% for six defects in a PCB dataset, which outperforms other existing models in terms of precision and recall. Compared with the original RT-DETR algorithm, the proposed method improves precision, recall, and mAP50 by 5.8%, 7.9% and 5.4%, respectively, In addition, the inference speed reaches 66.2 frames per second (FPS), which is deemed effective for the detection of small target defects in PCBs.
Wang et al. (Thu,) studied this question.