A-BEBLID: A Hybrid Image Registration Method for Lithium-Ion Battery Cover Screen Printing

Abstract

To address the problem of miss- and false-detection during quality inspection of lithium-ion battery cover screen printing (LBCSP), we propose a hybrid image registration method using a point-based feature extraction algorithm and nonlinear-scale space construction. Our proposed method integrates the AKAZE algorithm with the BEBLID descriptor, and is therefore named A-BEBLID. Facing the challenge of the inevitable offset caused by machine vibration during production, we combine a nonlinear diffusion filter with a local image descriptor to extract features from images, and then use the GMS algorithm to remove the wrong matching pairs. We tested the method on a dataset we created using images taken from actual lithium-ion battery production lines, named LBCSP. We also evaluated the method on the public HPatches dataset. The average precision achieved by A-BEBLID on the LBCSP dataset is 89% (threshold: 2 pixels), with a localization error of 1.11 pixels, while on the HPatches dataset, the average precision is 73% (threshold: 2 pixels), with a localization error of 1.52 pixels. Comprehensive experimental results also showed that the proposed A-BEBLID can outperform other approaches being compared to. The method can be further applied to other industry scenarios with similar image registration requirements.