PPR-Net: Patch-based multi-scale pyramid registration network for defect detection of printed labels
Detecting defects in printed labels is essential to ensure product quality. Reference-based comparison is a potential method to challenge this task, which is widely used for defect detection. However, this method gets poor performance under large deformation, due to the lack of ability of registering the testing image with the reference image. Therefore, accurate image registration is an urgent case for defect detection of printed labels. In this paper, a patch-based multi-scale pyramid registration network (PPR-Net) is proposed. First, an image patch splitting and stitching strategy is proposed, which is scalable in image resolution. Second, a multi-scale pyramid registration module is designed to fuse multiple convolutional features to enhance the registration capability for large deformation, which gradually refines multi-scale deformation fields in a coarse-to-fine manner. Third, a distortion loss function is introduced to improve text distortions of registered images. Finally, a synthetic database is generated based on real printed labels, to simulate defective printed labels with large deformation for performance comparison. Extensive experimental results show that our method dramatically outperforms other comparable approaches.