Object detection of welding defects in SMT electronics production based on deep learning

Nowadays, electronic products as an indispensable part of people's working life, its production quality requirements can no longer be ignored. From the once inefficient development of manual visual inspection to today's AOI equipment detection stage by machine substitution, the detection of circuit welding defects reflected the continuous innovation and development of science and technology. With the rapid increase in computer computing power in recent years, deep learning algorithms have risen again, which were neglected for a long time. In this paper, we study the neural network algorithm based on computer vision (CV) for the defects that often occur in PCBA The dataset used in this paper comes from data stored in the SMT line's AOI equipment provided by an electronics manufacturer. The first step to deal with the images is filtering. 759 pictures of SMT components was selected, which containing common defects, including insufficient solder, false welding and billboard. Then label these defects of the images with bounding box by the software labelimg. Finally, training models based on neural network with this dataset. Since traditional neural networks have limited ability to recognize special small targets such as solder joint images, we used a variety of convolutional neural networks (CNN) for classification and localization of defects. Various models such as Faster R-CNN, Yolov3, Yolov4, Yolov5 and Yolox were selected. The results of training are analyzed and compared, and the model with higher accuracy is selected for the detection of welding defect. It was concluded that Yolox-s had the highest recognition rate and was the best SMT welding defect detection model with mAP = 94.19%.