基于深度学习的多语种自然手写文本识别理论和方法

Driven by multiple requirements, such as industrial needs, human intellectual challenges and national strategy, multilingual unconstrained handwriting recognition has a broad and bright future. However, currently there exists few recognition theories and practices to support the inter-lingual, unconstrained handwriting recognition. This project studies the theories and methods for multilingual unconstrained handwriting recognition based on deep learning. We aim to properly expand current handwriting recognition theories. It is expected to push forward comprehensive handwriting recognition theories and extensively application of the proposed technology. The issues undertaken include: 1) Novel recurrent neuron unit and new deep network architecture, 2) Novel segmentation-recognition integrated framework supporting both online continuous writing and overlaid writing, 3) Segmentation-free framework based on deep recurrent neural networks, and 4) End-to-end handwriting recognition framework based on recurrent neural networks. ..Exploring the powerful representation learning in deep learning, the project is expected to propose novel handwriting recognition theories and methods, and solve the challenges in unconstrained Chinese handwriting recognition and multilingual handwriting recognition. Moreover, we plan to publish academic papers in prestigious journals and conferences. In addition, the project will develop unconstrained handwriting input method, camera-based handwriting recognition system and transcription system for multilingual handwritten essay.

多语种自然手写文本识别为来自产业界、人类智能极限探索以及国家战略需求上的多元因素所驱动，应用前景广阔。然而，目前能够支持用户自然书写习惯的跨语种文本识别理论和实践比较缺乏。本项目研究基于深度学习的多语种自然手写文本识别理论和方法，使用深度学习扩展现有文字识别理论，设计能够应用于多语种手写文本识别任务的多种新型框架和方法，促进文字识别理论和技术的全面发展及广泛应用。研究内容包括：新型递归神经单元和新型深度神经网络，支持联机双书写模式的切分-识别集成框架，基于深度递归神经网络的无切分识别框架和端到端手写文字识别框架等。将借助深度学习的高效特征表示能力提出一系列有效的手写文本识别新理论和方法，解决自然手写中文文本识别和多语种手写文本识别中的瓶颈；在国内外权威刊物和学术会议上发表一系列有影响的学术论文；研制联机自然手写输入法、脱机拍照手写文本识别原型系统和面向多语种手写作文识别系统。

多语种自然手写文本识别为来自产业界、人类智能极限探索以及国家战略需求上的多元因素所驱动，应用前景广阔。本项目面向用户终端、云计算、边缘设备的异构环境开展基于深度学习的多语种自然手写文本识别理论和方法研究，使用深度学习扩展现有文字识别理论，设计能够应用于多语种手写文本识别任务的多种新型框架和方法，促进文字识别理论和技术的全面发展及广泛应用。本项目研究了面向多语种文字识别任务的基础数据采集和自标注方法、多种架构设计理论和方法、新型目标函数的设计方法、面向移动端的运行优化技术等内容。本项目紧紧围绕多语种手写文字识别中的主要挑战问题，产出了一些典型理论成果和关键技术，在国内外权威刊物和学术会议上发表一系列有影响的学术论文，并在联机自然手写输入法、脱机拍照手写文本识别系统和面向多语种手写试卷识别系统中得以应用。