面向手写藏文古籍的跨字体文字识别方法研究

藏文古籍文献数字化是解决古籍“藏”与“用”的矛盾的有效手段，而藏文古籍文本识别技术是数字化的关键技术之一。尽管雕版印刷藏文古籍文本识别已经得到较好解决，但是手写藏文古籍因书写风格导致的不规则和多样化字体，因书写条件限制导致的复杂版面样式，以及因保存条件影响导致的图像退化等问题，准确识别其中的文字依然极具挑战。为此，本项目提出藏文古籍版面内容的多粒度表征模型，并据此基于多任务深度学习技术实现鲁棒的手写藏文古籍版面分析；提出基于图像生成和对抗学习的手写藏文古籍文本行图像修复方法，克服古籍文档图像退化、文字破损的问题；提出藏文字字形与字体特征的解耦模型和字丁级语义特征提取模型，据此实现跨字体的手写藏文字识别方法和针对开放和长尾类别的藏文字识别方法，提升处理相似字、生僻字和错别字的能力。本项目研究成果有望突破手写藏文古籍文本识别的技术瓶颈，为作为中华文化重要部分的藏文化的传承与发展提供技术支撑。

The digitization of Tibetan ancient books is an effective means to solve the contradiction between “preservation” and “utilization” of ancient books, and the text recognition technology of Tibetan ancient books is one of the key technologies of digitization. Although the text recognition of Tibetan ancient books in engraving printing has been well resolved, it is still very challenging to accurately recognize the text of handwritten Tibetan ancient books because of the irregular and diversified fonts from different writing styles, the complex layout styles caused by the limitation of writing conditions, the degrade images caused by the preservation conditions. To this end, the project proposes a multi-granular representation model for Tibetan ancient books, based on which we propose a robust layout analysis method for handwritten Tibetan ancient books using multi-task deep learning technology; to overcome the challenges from degraded image and incomplete text in ancient documents, we propose a restoration method for handwritten Tibetan ancient book images based on image generation and adversarial learning; we finally propose a decoupling model to extract font and glyph features separately, together with a character-level semantic feature extraction model for cross-font, open and long-tailed category based handwritten Tibetan character recognition for similar and unfamiliar characters and typos. This project is expected to provide solutions and technical support for handwritten Tibetan ancient book text recognition, which can further help the inheritance and development of Tibetan culture and the Chinese culture.