多层次集成的手写汉字识别系统
文献类型:学位论文
作者 | 肖旭红 |
学位类别 | 工学博士 |
答辩日期 | 1997-02-01 |
授予单位 | 中国科学院自动化研究所 |
授予地点 | 中国科学院自动化研究所 |
导师 | 戴汝为 |
关键词 | 综合集成 多层次集成 手写汉字识别 签名鉴定 联机 脱机 Metasynthesis hierarchically integrated pattern recognition Handwritten character recognition Signature verification Off-line On |
学位专业 | 模式识别与智能系统 |
中文摘要 | 本文的工作是在综合集成思想的指导下,围绕多层次集成型模式识别系统的 模型设计及各部分需要解决的一些重要问题展开的。以手写汉字的识别(包括联 机识别与脱机识别)与联机签名鉴定为实验背景,作者不仅对多层次集成型模式 识别系统的模型进行了研究与具体设计,而且对其中一些关键性的环节,如多分 类器的集成、特征的表达与匹配等提出了一些切实可行的新算法。 首先,介绍了目前集成型模式识别系统的研究目的与重点,分析了现有的机 器自动模式识别技术与人的模式识别间的差距及自动模式识别系统的主要不足 之处,指出模式识别系统需要各个层次上的综合与集成。 在上述基础上,提出了一个简单的多层次模式识别系统集成模型。这个模型 选择集成单元为单分类器与集成环节的信息交换点,通过它,各单分类器间及单 分类器与集成环节间可交换信息并互发请求,根据具体情况选择不同的特征及分 析方法。作为这一模型的一个实例,把有效特征分成普遍性特征与特殊性特征. 提供了一种将这两类特征结合起来,进行综合分析的方法;并把这一方法具体应 用于联机签名鉴定系统。 多分类器的集成是多层次集成系统的一个关键步骤。根据各分类器对不同类 别的分类能力的差异,提出了一个分类器线性集成模型,并将它用于脱机手丐汉 字识别的三个分类器的集成,取得了很好的效粜。 特征的表达与融合对模式识别系统起着很关键的作用、作者在这方面主要作 了两项工作:(1)在对统计特征法与结构特征法及将统计属性融人结构表达巾的 方法所存在的不足之处进行分析的基础上,提出了一种利用统计特征来表达汉字 结构的有效方法;(2).将静态特征(图象特征)与动态特征(书写的笔画、笔顺 等)结合起来,用于解决自由联机手写汉字的识别问题,取得了一定的效果. 对于结构基元的匹配这一在字符识别中应用广泛而又相当困难的问题,提出 了一种运算复杂度很低的准松弛匹配算法,为进一步进行汉字结构匹配方面的研 究打下了基础。在整体匹配策略方面,针对联机汉字,提出了一种“自上而下” 与“自下而上”相结合,以合适的子结构引导整个汉字结构匹配的策略。区别于 以往的子结构引导方法,通过采用字首及字尾子结构及子结构动态抽取与匹配的 策略,使算法的实现简单易行。 另外,在字符图象的预处理,如字符的分割、关键点的检测等方面也作了很 多有一定新意的工作。 在这些工作的基础上,作者独自完成了脱机手写汉字识别、联机手写 |
英文摘要 | Directed by the theory of metasystnesis,the emphasis of this research has been put on the theory study and model building of a hierarchically integrated pattern recognition system,as well as some key problems related to the system design.Experimenting on handwritten(including both Off—line and off-line cases)Chinese character recognition and on-line signature verification problems,the author has not only done research in the building and realization of hierarchical integration system models,but also proposed several feasible new approaches for some critical procedures,such as feature representation,structural matching,as well as the combination of multiple classifiers, etc.. Firstly,the author elaborates the main achievements on the research of integrated pattern recognition,analyzes the disparities between computer pattern recognition and human beings’recognition and points out the shortcomings of current pattern recognition system model.Based upon this,the concept of hierarchical integration is proposed and a simple multi—layer integrated pattern recognition model is provided.In the model,the integration units are chosen as the places for information exchange Through which,various classifiers,classifiers and integration units send requests and responses mutually,SO that different features and discriminating rules can be applied conditionally.As a practical example of the model,a classification approach based on the combination of the analysis of university and that of individuality is proposed:the extracted features are dynamically divided into two categories:universal features and individual features by learning from samples,these two kinds of features are adopted by different classifiers and procedures,their results are interactive through a priori knowledge and rules.This approach has been used in the on—line signature verification system. Combination of the results coming from different classifiers is a crucial step of the hierarchically integrated pattern recognition system.In light of that the discriminating abilities from classifier to classifier,and from class to class for a particular classifier due to the differences of classifier structures and that of the features classifiers use,a lineal‘ classifier integration model is proposed and applied to integrate the three classifiers for handwritten Chinese character recognition Feature representation and fusion plays an important role in a pattern recognition system.Two pieces of work have been accomplished in this aspect.firstly,a nOvel structural representation of Chinese characters by statistical values is proposed,on the basis of analyzing the shortcomings of previous representation approaches.Secondly, static features(features extracted from character image)and dynamic features(stroke order,stroke number,etc.)are combined to tackle the problem of on—line freely handwritten Chinese characte |
语种 | 中文 |
其他标识符 | 406 |
源URL | [http://ir.ia.ac.cn/handle/173211/5666] ![]() |
专题 | 毕业生_博士学位论文 |
推荐引用方式 GB/T 7714 | 肖旭红. 多层次集成的手写汉字识别系统[D]. 中国科学院自动化研究所. 中国科学院自动化研究所. 1997. |
入库方式: OAI收割
来源:自动化研究所
浏览0
下载0
收藏0
其他版本
除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。