Error Reduction in Handwriting Recognition with applications

减少应用程序手写识别中的错误

• 批准号: RGPIN-2014-04330
• 负责人: Suen, Ching
• 金额: $ 2.84万
• 依托单位: Concordia University
• 依托单位国家: 加拿大
• 项目类别: Discovery Grants Program - Individual
• 财政年份: 2016
• 资助国家: 加拿大
• 起止时间: 2016-01-01 至 2017-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=610968
• 关键词: Error; Reduction; Handwriting; Recognition; applications

The objective of this program is to conduct advanced innovative research in computer recognition of handwritten characters, words, and symbols. Handwriting is one of the most challenging subjects in the field of pattern recognition due to the infinite varieties of character shapes and qualities. Although handwriting recognition systems have the potential of reading bank cheques, envelopes, archival documents, utility payment slips, income tax returns, and other business forms, very few are actually used at the moment, due to the high error rate of the current recognition systems. Indeed, banks and different companies are spending billions of dollars each week to enter handwritten data manually into the computer. In the past, most research was focused on high recognition rates with less emphasis on the most difficult and costly problems of error rates. This research picks up this challenge by investigating different methods of reducing the error rate to increase the reliability of recognitions, to produce a new generation of recognition systems for full deployment in practical environments. This research aims to increase the intelligence and capability of computers so that they can read, at high speed and with great accuracy, the information written on various types of documents where billions of dollars are being spent each week to manually enter the handwritten data from documents into the computer. Based on decades of experience plus a world renowned reputation in (a) conducting intense research in handwriting recognition, (b) having developed numerous high performance recognition systems, (c) having guided more than 190 graduate students, post-doctoral fellows, and visiting academic/industrial scientists, and (d) playing a lead role in this field and constantly interacting with prominent researchers around the world, the applicant proposes the following program to minimize the error rate and maximize the recognition score: 1) To continue to investigate the causes of substitution errors and the drawbacks of current recognition systems, 2) To explore and evaluate a variety of configurations of error-reduction schemes with a cascade of uncertainty-removal modules, 3) To create very large databases and introduce new training techniques to separate good and bad samples, 4) To discover distinctive features and vital parts of individual characters derived from eye-tracking and perceptual studies. 5) To select dynamic sets of complementary features for various stages of multi-stage recognition systems to produce the most reliable and the best system to recognize handwritten data in different languages, English, French, Arabic, Chinese, etc. 6) To integrate the error-reduced recognition systems with the high-performance systems, and discover the critical paths of optimizing their overall performance and trade-offs, 7) To apply the results to conduct large-scale experiments on reading different kinds of real-world business documents, word spotting, e.g. custom's declaration forms, taxation forms, bank cheques, utility bills, and archival documents of historical value, 8) To Introduce this new breed of hybrid classifiers which can identify confusing character/word shapes, minimize costly substitution errors and maximize performance, so that the newly developed systems can be used in practice to save billions of dollars and a huge amount of manpower. 9) Expand this project to recognize other types of patterns such as irises, faces, palmprints, and pedestrian patterns. .

该项目的目标是在手写字符、单词和符号的计算机识别方面进行先进的创新研究。由于字符形状和质量的多样性，手写是模式识别领域中最具挑战性的课题之一。尽管手写识别系统具有读取银行支票、信封、档案文件、公用事业付款单、所得税申报表和其他商业表格的潜力，但由于当前识别系统的错误率很高，目前实际使用的很少。事实上，银行和不同的公司每周花费数十亿美元将手写数据手动输入计算机。过去，大多数研究都集中在高识别率上，而较少关注最困难且成本高昂的错误率问题。本研究通过研究降低错误率的不同方法来应对这一挑战，以提高识别的可靠性，从而产生可在实际环境中全面部署的新一代识别系统。 这项研究旨在提高计算机的智能和能力，使其能够高速、高精度地读取各种类型文档上书写的信息，而每周要花费数十亿美元将文档中的手写数据手动输入计算机。基于数十年的经验以及世界知名的声誉 (a) 对手写识别进行深入研究， (b) 开发了许多高性能识别系统， (c) 指导过 190 多名研究生、博士后研究员和访问学术/工业科学家，以及 (d) 在该领域发挥主导作用并不断与世界各地的杰出研究人员互动， 申请人提出以下方案来最小化错误率并最大化识别分数： 1）继续调查替代错误的原因和当前识别系统的缺陷， 2）探索和评估具有级联的错误减少方案的各种配置 不确定性消除模块， 3）创建非常大的数据库并引入新的训练技术来分离好样本和坏样本， 4）通过眼球追踪和感知研究发现人物的鲜明特征和重要部分。 5) 为多级识别系统的各个阶段选择动态的互补特征集以产生 最可靠、最好的系统来识别不同语言的手写数据，英语、法语、阿拉伯语、 中文等 6）将错误减少的识别系统与高性能系统集成，并发现关键路径 优化其整体性能和权衡， 7）应用结果进行大规模实验，读取不同类型的真实商业文档， 单词识别，例如海关申报表、纳税表、银行支票、公用事业账单和档案文件 具有历史价值， 8) 引入这种新型混合分类器，它可以识别令人困惑的字符/单词形状，最大限度地降低成本 替换错误并最大化性能，以便新开发的系统可以在实践中使用以节省 数十亿美元和大量人力。 9) 扩展该项目以识别其他类型的图案，例如虹膜、面部、掌纹和行人图案。 。