Home  ›  News  ›  Content

Associate Professor PENG Liangrui’s Research Group Wins DAS 2016 Nakano Best Paper Award

Recently, Associate Professor PENG Liangrui and her research group from the Department of Electronic Engineering at Tsinghua University published a paper entitled "CNN based Transfer Learning for Historical Chinese Character Recognition" at the 12th International Workshop on Document Analysis System held in Greece (DAS 2016), and has won the DAS 2016 Nakano Best Paper Award. The first author, doctoral candidate TANG Yejun, attended the workshop and made an oral presentation.

DAS has been a biennial international workshop in Optical Character Recognition (OCR) research field organized by the International Association for Pattern Recognition since 1994. The Nakano Best Paper Award was established in memory of the late Prof. Yasuaki Nakano, who was the general chair of DAS 1998.

Historical Chinese Character Recognition is a technology to convert historical document images into full-text retrievable text files. Because of its technical challenges, the traditional pattern recognition and machine learning methods have not yet provided a sound solution. Convolutional Neural Network (CNN) is one of the emerging deep learning methods in machine learning research field.  Usually training a complex convolutional neural network needs millions of mass labeled sample data. However, it can hardly get satisfactory results by applying convolutional neural network directly to historical Chinese character recognition problem due to the lack of sufficient labeled training samples. The method proposed in the paper firstly uses the parameters of convolutional neural network trained by modern printed traditional Chinese character samples, then conducts transfer learning by using a few labeled historical Chinese character samples, and finally obtains the model adapted to the historical Chinese character recognition problem.  The paper also provides thorough analysis about the technical details, including the comparison with different model structures and parameters, and sample selection schemes. The achieved results will benefit the development of full-text digitization technologies for historical Chinese documents which will further promote the protection and utilization of the precious resources of Chinese cultural heritage.

DAS 2016 Nakano Best Paper Award certificate.

For the research work related to historical Chinese character recognition, it was supported by a joint-funded project between the National Natural Science Foundation of China (NSFC) and the French National Research Agency (French: Agence Nationale de la Recherche, ANR), collaborated with Professor Frank LeBourgeois from the University of Lyon, France, and National Library of China which provided historical Chinese document samples for this project; For the research work related to deep learning, it was supported by an international collaboration project sponsored by Toshiba Corporation. PENG Liangrui was the principal investigator on China side for these two projects. The authors of the paper include TANG Yejun, PENG Liangrui, XU Qian, WANG Yanwei from Tsinghua University, and FURUHATA Akio from Toshiba Corporation.

TANG Yejun met Professor George Nagy, who was the first researcher studying Chinese character recognition in the international academic community.

PENG Liangrui has been engaged in the research work on multi-lingual OCR and historical Chinese document recognition for more than 10 years. Previously, with her supervision, master candidate FENG Jixiong’s research paper entitled “Gaussian Process Style Transfer Mapping for Historical Chinese Character Recognition” won the best student paper award in the 22nd Document Recognition and Retrieval Conference (DRR 2015), in February 2015.  PENG Liangrui's research group belongs to the Laboratory for Intelligent Image and Document Information Processing (also called TH-OCR Lab) at the Department of Electronic Engineering.  The TH-OCR Lab was established in the 1980s. Under the leadership of the late Professor WU Youshou (member of Chinese Academy of Engineering) and Professor DING Xiaoqing, teachers and students have developed world-leading level innovative technologies in Chinese and multi-lingual OCR, multi-modal biometric identification and video surveillance. PENG Liangrui is also a member of the Research Center for Media and Big Data Cognitive Computing recently established at the Department of Electronic Engineering.  The director of the research center, Professor WANG Shengjin, pointed out that the innovation of the paper was the combination of the cutting-edge technology of deep learning with the difficult historical Chinese character recognition problem.