The Study Of Normalization For Online Handwritten Recognition

Handwritten character recognition has been more and more popular due to the development of the electronic techniques. However, the largest difficulty it faces is the huge difference that different people write. There are great differences in the order, the shape of strokes and the shape of whole charater even for the same character. These will affect the recognition result and the normalization method is thus brought up to solve this problem.The objects Offline Character Recognition deals with are usually saved as images which store the character information through gray value of the pixels. Therefore, the normalization of OCR is simplex. We just need to adjust the character shape by constructing re-positioning function to map the pixels to a standard plane. Online character recognition faces a timing characteristic handwriting sequence, in which the writing trace and the beginning and end of all strokes are recorded as a set of points. The normalization of online charater is more complex. It mainly focus on handling the stroke deformation and changes of stroke orders.Up to now, online character recognition researches can be categorized into two groups. In one group, it translates the online handwriting information into an image and uses the offline approach to classify. In the other group, it works on the timing information and recognizes the character by matching templates.In this paper, the above-mentioned two kinds of online recognition methods are combined. The offline shape warping technology is introduced to online processing after twice DDA smoothing, includig MCBA,LDPI,P2DCBA,P2DLDPI. After that, the normalizaed handwriting sequence will be corresponded with the template to normalize the stroke order. And then, we proposed a method to minimize the difference between two corresponding strokes. It normalizes the angle and location of the strokes through rotating and displacing.We perform experiments on HIT-OR3C, the results shows that after all these normalization steps the recognizing accuracy has been improved. Compared to recognize online character through offline method, online normlization and recognization method gains a higher accuracy. In addition, by comparing the experimental results on MNISTS, we analyze and summarize the impacts made by different nomalization method on gray features and gradient features.