Error Diffusion Halftone Classification using Contrastive Learning

Abstract

Error diffusion halftoning is one of the widely adopted technique in printers, to transform the gray-scale image into its approximate binary version. Further, the classification of halftones is very important to facilitate inverse halftoning, source printer identification, forensics analysis and other halftone processing tasks. Practically, majority of the printed documents are unlabeled and hence hard to train using supervised approach. This study exploits the advantage of self-supervised learning (SSL), in particular the simplified framework for contrastive learning of visual representation in learning best representation features for halftones. As the data augmentation play a critical role in SSL models, and this study focus on optimization of the existing augmentations and also added new random augmentation techniques to enhance the feature learning. In addition, different variants of ResNet backbone is tried to find the ideal case, and the error diffusion dataset is also generated for analysis. From detailed experiments, it has been found that the proposed method can perform consistent with supervised learning approach without large labelled data.