An End-to-End Preprocessor Based on Adversiarial Learning for Mongolian Historical Document OCR

• Published in: PRICAI 2019: Trends in Artificial Intelligence Vol. 11672; pp. 266 - 272
• Main Authors: Su, Xiangdong, Xu, Huali, Zhang, Yue, Kang, Yanke, Gao, Guanglai, Batusiren
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2019; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Adversiarial learning; Binarization; Denoising; OCR; Preprocessing
• ISBN: 3030298930; 9783030298937
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-29894-4_21

In Mongolian historical document recognition, preprocessing mainly involves image binarization and denoising. This is a challenging task and greatly effects the accuracy of the recognition result. Concerning the fact that image binarization and denoising are both image-to-image tasks, this paper proposes an end-to-end preprocessor for Mongolian historical document OCR. The preprocessor is trained in an adversarial learning fashion and deal with binarization and denoising simultaneously. The input of the preprocessor is the color image of Mongolian document images, and the output is the clean binary images which can be used for word recognition. The preprocessor was trained on a limited dataset and performed better than the combination of binarization and denoising methods used in earlier Mongolian historical document OCR systems.