Research On Algorithm For Unconstrained Handwritten Numeral String Segmentation And Recognition

Handwritten numeral character recognition is a branch of OCR and it also is the important problem of pattern recognition. It is concerned how to auto-recognize the handwritten Arabic numerals on the papers using electronic computer, which has the abroad application prospects in the areas such as post tap, statistic report forms and so on. At the same time, the numeral strings segmentation is the key step for recognition, so it also has the important research value. In this paper we mainly discuss general methods for unconstrained offline handwritten numeral strings segmentation and recognition.Handwritten character recognition has been extensively studied for many years and a number of techniques have been proposed. However, it is still a difficult task and the result is still far from practice, especially when there are many connected numerals in the image, because of the different writer and different writing style. This paper is focused on unconstrained handwritten numeral string segmentation and recognition problem. The main contents of the thesis are as follows.(a) In numeral string segmentation stage, based on listing the present classic methods, the author proposed a novel method based on convex and concave feature. The method basically overcomes the defects of thinning and path-searching, so it gets the higher segmentation rate and shorter time.(b) We do not adopt neural network, which is the popular method for numeral character recognition at present, but propose a new method that it composes sequence code and regular expression using the character's convex and concave feature. Since global features of characters, the method has higher recognition rate and the recognition time is raised one time as well.(c) Other than numerals, we expand the method into the small-set characters. Therefore, we specially have a test about the capital English letters recognition. And we describe the full method, which is how to recognize the letters based on the convex and concave feature. Compared with other methods, its recognition rate and speed are improved.(d) At last, we have finished an unconstrained offline handwritten numeral strings segmentation and recognition demo system, Whose object are NIST SD19 and SHNID. Besides, the method proposed in the thesis was in contrast to other methods, and acquired better result.