【作者】 居琰；

【导师】 袁祥辉；

【作者基本信息】 重庆大学 ， 仪器科学与技术， 2002， 博士

【摘要】 信息融合技术已经成功地应用于众多的研究领域，在模式识别领域也有巨大的应用价值。通过研究可以发现，对于复杂模式识别问题，如手写体汉字识别，可以说目前还没有一个简单的方法可以达到较高的识别率和可靠度，每一种方法都有各自的优点、缺陷和不同的适用范围，不同的特征和匹配方法之间具有一定的互补性。因此，研究如何将不同的方法有机地结合起来以充分发挥各自的优势，克服其缺陷，从而构成信息融合型的识别系统，就成为当前模式识别研究的一个主要方向。本文主要在基于信息融合结构的手写体汉字识别理论和应用方面进行了以下工作：1、手写体汉字变形问题是手写体汉字识别中的关键问题，归一化处理是直接从汉字点阵图像上矫正手写变形、减小属于同一类别的不同模式之间差异的途径。通过对目前文献中的几种归一化处理算法的深入研究，提出了一种新的非线性归一化方法。该方法在进行空间坐标变换归一化处理时，采取线密度填充的算法，使得笔划密度的描述更为合理，归一化后的汉字点阵中笔划的分布更加均匀。2、汉字特征直接反映着汉字形体整体或局部分布状况，良好的特征应该使同一种汉字的不同书写样本之间的差异性尽可能小,而在不同汉字之间的差异性尽可能大。根据手写体汉字的特点，对特征提取进行了分析研究，提出了改进方向线素特征、笔划分区矩特征、扩展周边笔划方向特征等几种新的手写体汉字特征，实验表明所提取的几种特征是有效的。3、研究了基于自适应特征融合及模块神经网络的手写体汉字识别。在传统的特征提取方法基础上, 给出了多特征融合的一种方法，多个特征通过广义K-L变换降维并融合，产生的新特征吸收了单个特征的对模式分类的优势。同时, 分类器采用多神经网络模块结构，将以往BP网络规模大小与问题复杂性的矛盾转化为系统规模大小与问题复杂性的矛盾。利用遗传算法同时进行特征选择及网络结构优化，以构成有利于分类的自适应特征空间，在此分类子空间内，分别对各类训练样本进行选择优化，从而最终实现了一个自适应特征模块网络结构。手写体汉字的识别试验验证了所给方法的有效性。<WP=6>4、通过对神经网络集成的理论分析，提出了一种多级神经网络结构的手写体汉字识别模型。第一级采用主分量分析神经网络，用于提取相关字符特征的主分量值，减小后级网络的运作规模。第二级为改进的GLVQ神经网络对PCA网络降维处理后的主分量特征对手写体汉字进行分类。GLVQ算法是从最优化一个目标函数而导出的，该算法构造新颖，为克服学习向量量化算法存在的问题提供了一个新的思路。分析了GLVQ算法的数学理论基础，完善了学习向量量化算法的理论，以及基于这些理论设计出更高效的分类器算法并把算法应用于有限集手写体汉字识别研究中。该算法不但可以有效提高系统的识别率，而且具有良好的泛化能力，这与传统的BP算法相比，有着明显的优势。5、提出了一种多神经网络融合结构并将之应用于有限集手写体汉字识别。为了提高多神经网络分类器融合的效果，采用了一种改进的证据理论融合方法，通过对大量样本的统计，获得有关每个分类器识别性能的先验知识，将其作为证据合成的依据。针对证据融合中计算复杂问题，推导了一种快速算法。在有限集汉字识别系统中的实验结果表明，不同的特征和分类器从不同的角度刻画了手写体汉字图像的本质，充分利用这些特征和分类器提供的信息，作出一个更为客观的决策是可行的。更多还原

【Abstract】 The information fusion technique has been widely and successfully applied to many fields, and is valuable especially to pattern recognition. It has been turned out that, for complicated pattern recognition problems such as recognizing handwritten characters, there exists no reliable and simple way to perform a high recognition rate. Every method has its advantages, limitations and applications, and is complementary for each other. So how to combine the advantages of various methods to develop an information fusion recognition system and overcome the disadvantages became one of the hottest themes of pattern recognition. The main works of this paper are listed as follows:1. To handle the deformations is one of the key problems of handwritten Chinese characters recognition. The goal of normalization is to correct the deformation of handwritten character images and reduce the differences of the character images, which belong to the same class. A new nonlinear normalization method was developed, after I carefully researched the existed normalization methods. This method is based on line density method, and the line density filling in algorithm is used to perform the coordinate transformation. This makes the description of the stroke density more reasonable, and the distribution of strokes in normalized character image more proportional. 2. The features of Chinese characters characterize the unitary and local shape of the characters. Under a set of well-selected features, the distances of character samples of same class ought to be small; at the same time, the distances of character samples of different classes ought to be prominently large. According to the characteristics of handwritten Chinese characters, I developed some new features such as improved directional element feature, stroke sub-area moment, extended circumjacent stroke directional element feature, etc. These features are proved to be effective in experiments.3. I studied the handwritten Chinese character recognition techniques based on adaptive information fusion and module neural networks. I provided a new method of multi-feature fusion on the basis of the<WP=8>traditional feature-extraction methods. The high dimension of feature vector can be reduced by generalized K-L transformation, and the new feature vector is produce with the predominance of the single features for classification. The classifier is a multi-module neural network, so the conflict between the scale of BP network and the complexity of recognition tasks can be transformed to the conflict between the scale of recognition system and the complexity of recognition tasks. In the meantime, an adaptive feature subspace was constructed by using genetic algorithm to select features and optimize the structure of the neural network. The subspace of classification was optimized by training samples too. Finally, an adaptive feature-generating neural network module was formed. The recognition experiments have verified the method.4. By analyzing the theory of neural network integration, I developed a multi-level neural network model for recognition handwritten Chinese characters. The PCA network is used as the first level to extract the principle component of the features of characters, and reduce the scale of the neural network of subsequent level. The second level is a GLVQ network. The function of the level is to classify the principle component feature vectors produced by the PCA network. The novel GLVQ algorithm was induced to optimize an object function, it provides a new way to overcome the exist problem of LVQ algorithms. I analyzed and improved the basic mathematic theory of GLVQ, designed a more efficient classification algorithm based on the theory, and applied the algorithm to handwritten Chinese character recognition. The algorithm improves the recognition rate evidently, and has well capability of generalization. It is obviously superior to traditional BP algorithms.5. A multi-neural network fusion structure was developed and was applied to recognize the limited set of更多还原