A Method of Defect Detection for Focal Hard Samples PCB Based on Extended FPN Model

Abstract

Suffering from the diversity, complexity, and miniaturization of printed circuit board (PCB) defects, traditional detection methods are difficult to detect. Despite object detection has made significant advances based on deep neural networks, it remains a challenge to focus on small objects. We address this challenge by allowing multiscale fusion. We introduce a PCB defect detection algorithm based on extended feature pyramid network model in this article. The backbone is constructed by part of ResNet-101, in order to accurately locate and identify small objects, this article constructs a feature layer, which integrates high-level semantic information and low-level geometric information. Based on feature pyramid networks (FPN) network structure, using <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$1\times1$ </tex-math></inline-formula> convolution lateral fusion of the previous semantic information, the fused features use <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$3\times3$ </tex-math></inline-formula> convolution to obtain the final feature layer. The problem that PCB defects are difficult to classify is considered, the focal loss function is introduced. To reduce over-fitting in the training process, the original data are enhanced using image clipping and rotation. Through the quantitative analysis on PCB defect dataset, these results are the best to be used in fused low-level feature layer for detection of the mean average precision (mAP). This is 96.2% on the public PCB dataset, which is surpassing the state-of-the-art methods.