A coarse-to-fine registration network based on affine transformation and multi-scale pyramid

It is essential to guarantee product quality by detecting defects in industrial printed labels. Defect detection based on reference comparison is a common method to achieve this task. However, this method yields poor performance under large deformations of printed labels, due to the lack of accurate registration ability between testing images and reference images. Therefore, it is an urgent task to achieve accurate image registration for printed label defect detection. In this paper, a coarse-to-fine end-to-end registration network, named APPR-Net, is proposed to deal with the large deformation. First, we adopt a strategy of image patch splitting and stitching to improve the scalability of image resolution. Second, we design a four-stream affine transformation module followed by a multi-scale pyramid registration network, where a coarse registration is obtained by the former module and then gradually refined by the later network in a coarse-to-fine manner. Third, we introduce a distortion loss function to constrain the text distortion of the warped image after image registration. Finally, to simulate printed labels with defects and large deformation, we build a synthetic database based on real-world industrial printed labels for performance comparison. The results demonstrate that the proposed APPR-Net significantly outperforms other compared methods.