Component Welding Defect Intelligent Detection Based on Deep Learning

Components welding defect detection of PCB (Printed Circuit Board) is indispensable in PCB manufacturing process, which is the quality guarantee of electronic products. At present, there are I three difficult problems about PCB component defect detection in factory. Firstly, there are far more normal samples than negative samples because of difficulty in collecting negative samples. Secondly, industrial assembly line requires real-time performance, so that the detection model should be as tiny as possible. Thirdly, the factory often requires a 100% recall rate and a very low false positive rate, but these two indicators are mutually constrained. To solve the first problem, we use data rebalancing strategy to balance the dataset. On the one hand, we increase the number of negative samples by data augmentation technologies. On the other hand, we cluster the positive samples and appropriately remove some positive sample from each cluster. To solve the second problem, we designed a lightweight network that introduced feature fusion strategies to improve accuracy of detection. To solve the third problem, we will design an inter-class decision mechanism to make a second decision in the output of network. Finally, our model achieves better mAP of 99.90% than other methods on our dataset, and a 100% defect recall and 97.15% defect precision are achieved on the dataset provided by ZTE.