Improved bare PCB defect detection approach based on deep feature learning

Abstract

Robust and precise defect detection is of great significance in the production of the high-quality printed circuit board (PCB). However, due to the complexity of PCB production environments, most previous works still utilise traditional image processing and matching algorithms to detect PCB defects. In this work, an improved bare PCB defect detection approach is proposed by learning deep discriminative features, which also greatly reduced the high requirement of a large dataset for the deep learning method. First, the authors extend an existing PCB defect dataset with some artificial defect data and affine transformations to increase the quantity and diversity of defect data. Then, a deep pre-trained convolutional neural network is employed to learn high-level discriminative features of defects. They fine-tune the base model on the extended dataset by freezing all the convolutional layers and training the top layers. Finally, the sliding window approach is adopted to further localise the defects. Extensive comparisons with three traditional shallow feature-based methods demonstrate that the proposed approach is more feasible and effective in PCB defect detection area.