【摘要】 传统的字符归一化方法只是对字符的大小进行改变,它与字符类别无关,也无法对字符的旋转、扭曲等变形进行矫正。本文采用一种与字符类别相关的自适应归一化方法——全局仿射变换,用于神经网络手写体字符识别中,利用已知类别的参考模板对输入字符进行全局归一化,同时对字符的旋转、扭曲等变形进行校正,归一化准则定义为参考模板与输入模板间的最近邻距离。并采用对权值的迭代算法得到最优的匹配模板。采用不同神经网络识别系统对该方法验证,表明该方法可使系统识别率得到较明显提高。更多还原

【Abstract】 Conventional normalization methods have only been limited for correcting position and size change of character. Those methods can not be used for processing the rotation, shearing and other kinds of shape distortion. In this Paper, we use an adaptive or category--dependent normalization method based on Global Affine Transformation (GAT) which normalizes an input pattern against each reference pattern. This method can be used to correct the rotation, shearing and other kinds of shape distortion that occur in handwriting. The rule of normalization is defined as the nearest distance between the reference patterns and the input patterns. Experimental results show that the correct identification rate for handwritten characters has been raised obviously by using this normalization method.更多还原