Character-interested binary-like image learning for text image demoiréing

Abstract

Despite the fact that the text image-based optical character recognition (OCR) methods have been applied to a wide range of applications, they do suffer from performance degradation when the image is contaminated with moiré patterns for the sake of interference between the display screen and the camera. To tackle this problem, we propose a novel network for text image demoiréing. Specifically, to encourage our study on text images, we collected a dataset including a number of pairs of images with/without moiré patterns, which is specific for text image demoiréing. In addition, due to the statistical differences among various channels on moiré patterns, a multi-channel strategy is proposed, which roughly extracts the information associated with moiré patterns and subsequently contributes to moiré removal. In addition, our purpose on the text image is to increase the OCR accuracy, while other background pixels are insignificant. Instead of restoring all pixels like those in natural images, a character attention module is conducted, allowing the network to pay more attention on the optical character-associated pixels and also achieving a consistent image style. As a result from this method, characters can be more easily detected and more accurately recognized. Dramatic experimental results on our conducted dataset demonstrate the significance of our study and the superiority of our proposed method compared with state-of-the-art image restoration approaches. Specifically, the metrics of recall and F1-measure on recognition are increased from 56.32%/70.18% to 85.34%/89.36%.