The Research On Off-line Handwritten Chinese Character Recognition

Chinese characters are the results of Chinese cultural accumulation, which has a long history, and they represent Chinese people's wisdom. Along with the progress of technology and the development of information age, more and more people in the world begin to use the Chinese character, so automatic Chinese character recognition is being paid more attention by the researcher in the pattern recognition domain. Usually, Chinese character recognition includes printed Chinese character recognition and handwritten Chinese character recognition. By far, printed Chinese character recognition has been used in many areas, and had a tendency toward to higher performance and more perfect user. However, Handwritten Chinese character recognition is difficult at all times in the pattern recognition area and still in the state of experimental trial. This paper studies handwritten Chinese character recognition, which includes feature selection and feature extraction.Feature selection is an important technology for dealing with the original feature, aiming for removing useless,redundant or irrelevant features. This paper proposes a new feature selection algorithm named MFOS based on FOS. The new algorithm selects the features which represent the most features in the datasets, forms effective feature subsets to represent one Chinese character. It extracts important feature information of the Chinese character, at the same time, reduces unnecessary relation among features in order to classify and recognition effectively.Feature extraction is a very important step for Chinese character recognition. In last 20 years, many Chinese people and foreigners have put forward a lot of methods of feature extraction aiming to this problem. Features are divided into two kinds: statistic feature and structure feature. Statistic feature contains character context contour feature, stroke direction feature and so on, structure feature includes feature point, stroke segment, stroke etc. According to the feature extraction of the handwritten Chinese character recognition, an improved extracting stroke planes method is proposed. This paper brings forward the elastic stroke length to extract the stroke planes based on normalized images. The advantage is whether or not to extract the black point on the scanning beam is decided by the numbers of the black points on the scanning beams from different directions. This method catches hold of every image's imperceptible information, and improves the recognition rate definitely. Theory analysis and experiment results indicate the effectiveness of this method.Based on stroke plane extraction and dynamic meshing partition, a new method for Chinese character feature extraction is proposed. Considering overlapped dynamic meshing can overcome the sensitivity caused by the stroke displacement and local distortion, and fuzzy membership show the importance of points which make up of the Chinese character image, so the basic idea is to divide the Chinese character image into four different stroke planes by dynamic meshing, then allocate the fuzzy membership for each meshing, finally, compute the weighted accumulation histogram aimed for obtaining the feature for each meshing. The new method conquers the unstable problem caused by strokes'distortion. The experiment results on NUST603HW handwritten Chinese character database of Nanjing University of Science and Technology verify the effectiveness of the proposed method.