PCB defects target detection combining multi-scale and attention mechanism

The detection of PCB defect quality plays an important role in PCB fabrication. However, the size of the PCB defects is too small to identify. In order to improve the detection efficiency of existing algorithms, a joint multiscale PCB defect target detection and attention mechanism, which named RAR-SSD, was proposed. By using lightweight receptive field block module (RFB-s) with an attention mechanism module, we built a wider range of effective focused features, which exploited the importance of different features in different channels without increasing the computing power of the network. In addition, we built a feature fusion module to efficiently fuse low-level feature information with high-level feature information to produce a more complete feature map and improve the accuracy of fault recognition. The proposed network improved the fault recognition accuracy of PCBs by 2.23% over the original SSD algorithm, with a recall rate of 6.51% and an F1 value of 4.85%, the model has greatly improved in terms of detection performance. The optimized algorithm has significant speed and accuracy advantages over the algorithms YOLOv3 and YOLOv5. Experimental results show that the proposed RAR-SSD model has good performance in detecting small and medium size targets for defects in the PCB manufacturing process and is of some guidance for the subsequent detection of PCB defects.

Automated summary

In order to improve the detection efficiency of PCB defects, a joint multiscale PCB defect target detection and attention mechanism, named RAR-SSD, was proposed. By using the lightweight receptive field block module (RFB-s) with an attention mechanism module, the proposed network built a wider range of effective focused features and efficiently fused low-level feature information with high-level feature information. The optimized algorithm improved the fault recognition accuracy of PCBs by 2.23% over the original SSD algorithm, with a recall rate of 6.51% and an F1 value of 4.85%, showing significant improvements in detection performance and outperforming YOLOv3 and YOLOv5 algorithms.