Abstract

AbstractIn printed circuit board (PCB) defect detection, there are some imbalances, which generally consists in feature level and sample level. The feature imbalance problem is caused by the semantic loss of features in the feature maps. The sample imbalance is caused by two aspects, defects of different scales and the incomplete utilization of negative sample information. These can affect the accuracy of defect detection. Resolving these imbalances is an important research direction to improve the accuracy of detection. To alleviate the impact of imbalance, a sample and feature equalization-based R-CNN detection algorithm (Feature and Simple Balance R-CNN, FaSB R-CNN) is proposed. It contains three innovative components: Feature Balance Processing (Feature Balance Algorithm and FPN), GA-RPN and IoU-Balanced Sampling. Using the above three components in the basic Faster R-CNN system, our method (FaSB R-CNN) has a large improvement in PCB defect detection. FaSB R-CNN achieves 8.18% higher precision on PCB Data set than the Faster R-CNN. Compared with other object detection algorithms, the FaSB R-CNN can also achieve relatively good results and has good practical engineering application value.KeywordsDefect detectionImbalancePCBDeep learning

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call