【作者】 温昌兵；

【导师】 杨扬；

【作者基本信息】 北京科技大学 ， 系统工程， 2005， 硕士

【摘要】 在金融保险、邮政税务、医疗卫生等各个领域,都需要将手写汉字转变为计算机内部可以保存的形式,以便能方便信息的交流,目前这种需求正逐年增加。因此对脱机手写体汉字识别技术的研究有广阔的应用前景。同时脱机手写体汉字识别一直是模式识别研究领域中的难点,囊括了模式识别领域中的所有典型问题,如特征选择、分类器选择以及样本集选择,因此对于它的研究具有深刻的理论意义。本课题针对脱机手写体汉字进行特征提取技术的研究,并设计了一个实验性的脱机手写体汉字识别系统。本文提出一种新颖的将局部特征与全局特征进行融合的特征提取方法。其中局部特征是基于结合Gabor滤波器的弹性网格技术进行提取的,而全局特征是基于Zernike矩变换进行提取。Gabor滤波器对笔划的方向具有很强的选择性,通过设计良好的Gabor滤波器组可以提取汉字在“横、竖、撇、捺”四个方向的笔划。在得到汉字四个方向的笔划分解子图像之后,利用弹性网格技术在原图像上构造弹性网格,将弹性网格作用于分解之后的汉字子图像上,然后统计各子图像在网格内黑像素点的概率分布作为局部特征。Zernike矩能很好地描述汉字的全局特征,它具有位移、旋转和变换不变性。两种特征融合后能互相弥补不足,具有更强的分类能力。最后,基于距离分类器对测试样本进行分类识别,验证文中提出的特征提取方法。实验结果表明,本文提出的特征融合方法的识别率达到92%以上,高于单独使用其中任一种特征进行识别的识别率,证明了特征融合方法的有效性。更多还原

【Abstract】 In the fields such as finance, insurance, post and revenue, we need to transformhandwritten Chinese characters into digital information which can be saved in computers so thatthey could be intercommunicated conveniently. And the demand is increasing year by year. Soresearch on off-line handwritten Chinese character has a broad application background.Meanwhile, off-line handwritten Chinese character recognition is a difficulty in theresearch field of pattern recognition, which covers all typical problems in pattern recognitionsuch as feature extraction, classifier selection and sample collection. So there is academic valuein research on it.This paper aims at research of feature extraction on handwritten Chinese characters, andbesides, an experimental off-line handwritten Chinese character recognition system is designedin this paper. In this paper, a novel extraction method is applied, which involved feature fusionof local feature and whole feature. The local feature is extracted based on Gabor transformationand elastic meshes, and the whole feature is extracted based on Zernike moment transformation.Gabor filters have the strong ability of distinguishing different directions. First we can design aset of filters to decompose an image of a Chinese character into four different directional strokesegments, namely horizontal, vertical, right-diagonal, and left-diagonal. After that, elasticmeshes are built from the original image and applied in the four stroke images. Then the pixelsdistribution of every mesh in every image could be calculated and characterized as the localfeature. Being motion, rotation and transformation invariant, Zernike moments can extract thewhole feature of Chinese characters. The fused new features have stronger class separabilitypower because they can offset the disadvantages of the two former ones.The experimental results show that the recognition rate of this method is greatly improvedto over92%, which indicate that the method is effective.更多还原