SELF-SUPERVISED LEARNING FOR SCANNED HALFTONE CLASSIFICATION WITH NOVEL AUGMENTATION TECHNIQUES

The current halftone classification models use supervised learning, which requires a large dataset of labeled images. However, in practical situations, the source halftone type is often unknown, making it difficult to create such a dataset. These models are typically trained on synthetic halftone images and perform poorly on scanned halftone images. To address this issue, a new self-supervised learning (SSL) model has been proposed, based on Barlow Twins (BT) and Blue Noise (BN) dithering. In addition, an effective patch swapping augmentation technique has been developed to improve accuracy and speed up the training process. The pre-trained model is then fine-tuned using a modified progressive multi-granularity model with limited halftone labels. In overall, the proposed model outperforms existing halftone classification algorithms, becoming the state-of-the-art method.