Research On Solder Joint Defect And Component Detection Method Based On Deep Learning

Circuit board solder joint defect detection and component detection are critical to the quality of electronic product hardware production.Traditional automatic optical inspection equipment requires customized registration and design procedures for each circuit board.The steps are complicated and time-consuming.The judgment standard is relatively simple and rigid.Poor lighting and changes in the color of the circuit board will affect the accuracy of detection.It is not cost-effective for production scenarios with many varieties and small batches.In recent years,with the rapid development of computing resources and the reduction of hardware costs,deep learning has risen again.It has shown excellent performance in the field of object detection,speech recognition and other research fields,and has been widely used.Therefore,it is of great research significance and industrial application value to introduce the deep learning method into the research of circuit board solder joint defects detection and component detection.This article uses two classic object detection algorithms and proposes an improved algorithm,the specific contents are:(1)The classical Faster R-CNN algorithm and YOLO V3 algorithm are utilized to detect solder joint defects and component of printed circuit board assembly(PCBA).Both are end-to-end object detection algorithms.The former uses VGG16 deep neural network as the feature extractor,realizes candidate bounding box recommendation through region proposal network(RPN),classifies and locates finally,which belongs to two-stage algorithm of object detection algorithm.The latter uses Darknet53 for feature extraction,then classify and locate objects,which belongs to one-stage algorithm of object detection algorithm.The experimental results show that they have high accuracy,but the mean average precision(mAP)of Faster R-CNN algorithm based on VGG16 is significantly lower than that of YOLO V3.(2)An improved Faster R-CNN algorithm is proposed.The first eight stages of EfficientNet B7 network structure in EfficientNet series are used to replace VGG16 for feature extraction.The generalized intersection over union(GIoU)is used as the loss of bounding box regression,which avoids the problem of minimizing the loss of ln and maximizing the intersection over union without strong correlation in some cases.The Swish function with better performance is used to replace the ReLU function as the activation function.The experimental results show that the improved algorithm has a significant improvement in accuracy,and the mAP is close to 0.99,which is of great significance for further industrial application research.