Abstract
Circuits on different layers in a printed circuit board (PCB) must be aligned according to high-precision fiducial mark images during exposure processing. However, processing quality depends on the detection accuracy of fiducial marks. Precise segmentation of fiducial marks from images can significantly improve detection accuracy. Due to the complex background of PCB images, there are significant challenges in the segmentation and detection of fiducial mark images. In this paper, the mARU-Net is proposed for the image segmentation of fiducial marks with complex backgrounds to improve detection accuracy. Compared with some typical segmentation methods in customized datasets of fiducial marks, the mARU-Net demonstrates good segmentation accuracy. Experimental research shows that, compared with the original U-Net, the segmentation accuracy of the mARU-Net is improved by 3.015%, while the number of parameters and training times are not increased significantly. Furthermore, the centroid method is used to detect circles in segmentation results, and the deviation is kept within 30 microns, with higher detection efficiency. The detection accuracy of fiducial mark images meets the accuracy requirements of PCB production.