基于全类分类器与子集分类器融合的脱机手写汉字识别研究
文献类型：学位论文
| 作者 | 高天孚 |
| 学位类别 | 工学博士 |
| 答辩日期 | 2009-06-03 |
| 授予单位 | 中国科学院研究生院 |
| 授予地点 | 中国科学院自动化研究所 |
| 导师 | 刘成林 |
| 关键词 | 脱机手写汉字识别 分类器融合 相似字判别 复合距离 D-S证据理论 handwritten Chinese character recognition classifier combination compound distance similar character recognition D-S evidence theory |
| 其他题名 | Handwritten Chinese Character Recognition by Fusing Classifiers of Different Category Sets |
| 学位专业 | 模式识别与智能系统 |
| 中文摘要 | 随着模式识别方法的逐渐成熟和个人计算机性能的迅速提高，手写汉字识别技术取得了很大进展，但现有方法对脱机手写汉字识别的性能还存在明显不足。本文针对大类别集手写汉字识别的分类器设计进行研究，在基于全类分类器与子集分类器融合方面做了一些工作。主要研究工作如下: 一、提出了基于线性鉴别分析（LDA）的复合距离方法，该方法能够有效的区分汉字中的相似字，并取得了很高的整体识别率。我们证明了在一些严格的假设条件下，以往学者提出的复合马氏距离是我们提出方法的特例。虽然我们提出的方法需要一些额外的存储量，但当在原始特征空间计算相似对的LDA鉴别矢量的时候，我们提出的方法能够获得比MQDF和复合马氏距离高得多的识别率。 二、比较分析了线性鉴别分析（LDA）、异方差鉴别分析（HLDA）、近似信息鉴别分析（AIDA）和线性支持向量机（SVM）在相似字鉴别方面的性能，实验结果表明绝大多数相似字是线性可分的。HLDA和AIDA与LDA相比没有明显差别，这说明即使是非常相近的汉字，它们的类别中心也离得足够远。线性SVM和LDA相比在区分相似字上也没有明显区别。 三、提出了全类分类器与不完全两类分类器融合的一个概率性框架。全类分类器与两类分类器的输出被分别转化为所属类别的概率，然后全类分类器与相关两类分类器通过多数投票法、筛选法、对耦合（Pairwise coupling）、最小最大方法、纠错输出编码、排除解码方法进行融合。这些方法考虑到了全类分类器的多个候选，最后的识别结果比基于LDA的复合距离方法有了进一步提高。尤其在全类分类器识别率不高的情况下，效果更加明显。 四、提出了一种基于Dempster-Shafer（D-S）证据理论的融合大类别全类分类器（基分类器）与子集分类器的方法。通过引入虚拟分类器方法（子集内类别个数大于2）和否定概率法（子集内类别个数是2）来估计子集外类别概率。实验结果表明，该方法能明显地提高基分类器的识别率。 本文提出的一些算法在ETL9B和CASIA手写汉字样本数据库上测试，得到的识别率达到了世界先进水平。 |
| 英文摘要 | With the development of pattern recognition methods and computing technology, the performance of handwritten Chinese character recognition (HCCR) has been improved significantly. However, the recognition accuracy of unconstrained handwritten characters is still not satisfying. With the aim of improving the classification accuracy of large category set HCCR, we study on the combination of all-class classifier and subset classifier, and have achieved the following results. First, we propose a linear discriminant analysis (LDA)-based compound distance method for large-category set classification of HCCR, which can effectively discriminate similar characters and largely improve the overall classification accuracy. The LDA-based method is an extension of the previous compound Mahalanobis function (CMF). We prove that under restrictive assumptions, the previous CMF is a special case of our LDA-based method. We evaluated the methods in experiments on the ETL9B and CASIA databases using the modified quadratic discriminant function (MQDF) as baseline classifier. The results demonstrate the superiority of LDA-based method over the CMF and the superiority of discriminant vector learning from high-dimensional feature spaces. Compared to the MQDF, the proposed method reduces the error rate remarkably. For linear discrimination of similar characters, we compare the performance of LDA, heteroscedastic linear discriminant analysis (HLDA), approximate information discriminant analysis (AIDA) and support vector machine with linear kernels (Linear SVM). Experiments on the CASIA database show that all the four methods yield comparably high accuracies. This indicates that most similar handwritten Chinese characters are linearly separable and LDA is sufficient to perform this task. For combining a baseline all-class classifier and incomplete pair discriminators, we then propose a unified probabilistic framework. Under this framework, the outputs of the baseline classifier and pair discriminators are transformed to two-class probabilities, which are then fused by pairwise coupling (PWC), filter method, error correcting output codes (ECOC), and some other methods, to make the final decision. This framework considers more than 2 candidates given by the baseline all-class classifier and performs better than the LDA-based compound distance method. Finally, we propose a solution for combining a baseline all-class classifier and subset classifiers based on the Demps... |
| 语种 | 中文 |
| 其他标识符 | 200518014628060 |
| 源URL | [http://ir.ia.ac.cn/handle/173211/6215]  |
| 专题 | 毕业生_博士学位论文 |
| 推荐引用方式 GB/T 7714 | 高天孚. 基于全类分类器与子集分类器融合的脱机手写汉字识别研究[D]. 中国科学院自动化研究所. 中国科学院研究生院. 2009. |
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。