Recurrent spatial transformer network for high‐accuracy image registration in moving PCB defect detection

Abstract

Defects detection is an extremely important part in the production of printed circuit board (PCB) to guarantee the quality and reliability. A widely used and researched method for this task is referential comparison method. However, this method gets poor performance on the situation that PCB moves on the conveyor, due to the lack of ability to register test images with referential images. Therefore, accurate image registration which estimates the affine transformation between the captured test images and the referential images becomes an urgent problem to be solved. In this study, a spatial transformer network is proposed. In particular, a recurrent image registration strategy is introduced to step-by-step register the images in a recurrent progress. Furthermore, a referential part is designed to help training the network. Besides, in order to simulate the real moving PCB images and train the proposed model, a factitious dataset is generated by applying random affine transformations to real PCB images. Experiments show that recurrent spatial transformer network can achieve pixel-level accurate image registration. The defects detection precision of referential comparison method has a great improvement by using the registration algorithm.