【摘要】 文中提出了一种新的手写体数字识别的结构方法，将字符轮廓划分为特征片段：凸弧、凹弧、直线段、端点、洞，并由特征片段得到特征基元，构成对字符结构的完整描述．文中还提出了一种轮廓点的曲率滤波方法以及利用字符的特征基元序列测量两样本之间的相似度方法．在此基础上，提出了一个两阶段的识别方法及两种拒识准则．在第一阶段，通过基于相似度比较的分类器，对每个待识字符识别或拒识．拒识的字符进入第二阶段识别．拒识分两种情况：待识字符与任一类训练样本都不相似，称为第一类拒识；与两类训练样本都相似，为第二类拒识．在第二阶段，对于第一类拒识，重新提取特征后再入分类器．对于第二类拒识，用一个专用的二义识别过程识别．实验结果表明该方法的总体性能比前人的方法有所提高更多还原

【Abstract】 This paper proposes a new structural method for the recognition of handwritten numerals. Feature cells, which are derived from the primitive segments such as convex arcs, concave arcs, line segments, end point and holes from the contours of numeric characters, are used to describe numeral characters. A method of filtering the curvatures of contour points is developed. A method to measure the similarity between two samples based on the sequences of feature cells is proposed. A two stage recognition methodology is also presented, in which two rejection criteria are introduced. In the first stage of recognition, an input sample is given an identity or categorized as either first class or second class rejection based on similarity measures between the input sample and each of the ten numeral classes. In the second stage of recognition, strategies are adopted to modify the structural description of the input sample if it is in first class rejection and a classifier focussed n pairwise discrimination is applied if the input samples is in second class rejection. Experimental results indicate that the overall performance of the proposed method compares favorably with those achieved by other methods found in the literature.更多还原