Abstract: Accurate detection of tiny defects in printed circuit board processing is the prerequisite to ensure the quality of electronic products. Due to the small feature size and complex circuit layouts, the existing target detection methods have many shortcomings. To solve this problem, a YOLOv3-dense target detection model is proposed for printed circuit board defect detection based on YOLOv3 algorithm. First, dense connection network modules are used to replace some residual units in the feature extraction network so as to enhance feature reuse of the network. Then, the loss function is improved, and the generalized intersection ratio between the prediction box and the true value is used to solve the problem that the optimization cannot continue when the intersection ratio is zero. The experimental results show that compared with other models, the proposed model can improve the recognition accuracy and reduce the model size.

Key words: small target detection,  YOLOv3 algorithm, densely connected convolutional networks, printed circuit board defects