摘要
给出了一个高性能的自由手写数字识别系统，提出了在不同尺度下抽取结构上的粗特征和细特征并分层实现的方法。在粗分类阶段，基于结构粗特征用分类树进行稳定的粗分类，并进行属性确认，还用小波分解技术进行细分类识别；又提出并实现以微小线段及其邻域作为比较区域的相关匹配方法来减少候选类别数（限制性分类），最后，提出并实现用二分类ＢＰ网对非相似二类别字进行非限制性分类，另外，在识别的前端，还对书写工整的文字进行Ｆｕｚｚｙ描述分类，这一系列的分级识别带来了系统的高速度、高识别率。
In this paper a combinational method for recognition of unconstrained handwritten digits is proposed. To solve variation problem, three groups of feature at three different scales are extracted. Structure features are used to roughly describe the input numeral pattern, on the basis of which, attribute reasoning and identifications are implemented by means of tree classifier in the first stage, and DWT and a modified matching are carried out in the second stage. At last stage, the dicthrition structure of single layered BP network are used for the discrimination of dissimilar patterns in the same catagory. About twenty percent of well written digits are recognised by Fuzzy classifier at the beginning of the first stage. Experimental results show that the proposed recognition method is of better performace.
出处
《高技术通讯》
EI
CAS
CSCD
1998年第3期25-28,共4页
Chinese High Technology Letters
基金
北京大学视觉听觉信息处理国家重点实验室资助项目
关键词
文字识别
自适应预处理
模式识别
Digit recognition, Structure feature, Attribute reasoning, Discrimination
