FPC surface defect detection based on improved Faster R-CNN with decoupled RPN
Weiru Luo, Jiaxiang Luo, Zhiyu Yang
Abstract
The surface defects of FPC are characterized by large scale variation and shape variation. How to quickly detect the surface defects of FPC has become a challenge in the quality control of the electronic manufacturing process. In this paper, we first develop an FPC surface defect method based on the Faster R-CNN object detection model. Based on the method, the FPN multi-scale feature fusion structure is then introduced, and the multiple receptive field fusion module (MRFM) is proposed to improve the ability of the model to extract large scale and multi-scale features. Finally, by embedding the multiple receptive field fusion module in the FPN multiscale output, we reduce the coupling between RPN and BBox head and improve the detection performance. To evaluate model performance, we build a dataset of FPC surface defects containing seven general defect types: short, open, pinhole, line damage, broken hole, exposed copper, and scratch. The experimental results show that the proposed model can achieve the mAP of 0.9557 and the mean recall of 0.9699 in the FPC surface defect dataset, which is better than the existing algorithmic model.