Defect Detection of Printed Circuit Boards Using EfficientDet

Abstract

Printed circuit board is an important part of electronic industry. The detection of PCB defects could improve the reliability of PCB and electronic devices. Comparing to the traditional inspection methods, the methods based on Convolutional Neural Network have higher accuracy, faster speed, and more robustness. In this paper, a PCB defects inspection method based on EfficientDet-D1 are proposed, which uses EfficientNet-B1 as the feature extraction backbone. Besides, we use K-means clustering to predict more accurate anchor ratios. The results of our work show that the method achieves a similar mAP with other PCB defect detection methods while costing less time in detection.