手写数字识别实验系统的研究

以无限制手写数字的机器自动识别系统的研究对象，旨在达到对书写人员、笔墨、纸张、字体大小、取向、位置不作限制，对常见书写风格均能识别。

RESEARCH ON THE EXPERIMENTAL SYSTEM FOR RECOGNITION OF HANDWRITTEN DIGITS

This thesis deals with the recognition of handwritten digits generated by different writers, aiming at no constraints on paper, ink, pen, size and position of digits. During the stage of feature extraction we use Fuzzy technique together with some topological and a few of geometrical features which are easily extracted can help us to obtain fast judgements for some of well written digits. We extract the macro-feature to roughly describe the input numeral pattern, and implenent this process at the different scales. And we continue feature extraction to the local features as well. We also make use of micro-information extracted from the small region or shorter stroke or some pixel in the stroke. Reasoning policy in our work is based on experienced knowledge which is expressed by about 600 items of rule. In addition to reasoning, attribute verification makes the result of preclassifying few errors. In our experimental system we strengthened discrimination of digits attributed to different catagories which are in the same subsets by means of the application of DWT to the discrimination of dissimilar stroke for similar digit pattern, matching by distance for similar digit patterns and the application of BP network to the dichotomy in the subsets. Experimental results show that our experimental recognition system has potentiality to learn to some extent.