The Research Of Multi-Classifier Ensemble Based On Offline Handwritten Chinese Recognition

Chinese character is the most gorgeous treasure of Chinese culture which with thousands of year's history. It makes indelible contributions to the development of Chinese culture. In the current high-speed development information age, how to contact with the computer more efficiently has become a very important bottleneck of the interaction between human and machine. Making the computer to recognize Chinese characters automatically is the most direct and pressing issue of Chinese information processing. Chinese character recognition is an important branch of pattern recognition, also is the most difficult issue of character recognition field. It involves many disciplines, such as pattern recognition, image processing, digital signal processing, natural language understanding, artificial intelligence, fuzzy mathematics, information theory, Chinese information processing and so on. As an integrated technology, it has an important theoretical and practical value in the fields of Chinese information processing, office automation, machine translation, and artificial intelligence.This paper presents firstly the basic processes of off-line handwritten character recognition, and introduces various theories and methods used in each aspect of off-line handwritten character recognition processes. Then makes a summary of various classifiers and ensemble methods commonly used in the off-line handwritten Chinese characters recognition.After comprehending and analyzing the various classifiers and integration of multi-classifiers, a new method of multi-classifier ensemble is presented in this paper. The method use two-level classifiers to ensemble. An improved Euclid distance classifier in the first level is used to make a roughing set for the test samples, and then the one-against-rest SVM classifiers are applied to recognize the handwritten Chinese more exactly. Last the two-level classifiers are ensemble by Bayes law.In addition, in the preprocessing stage of the recognition of non-specific handwritten Chinese manuscripts, this paper presents a simple, rapid segmentation method of Chinese characters. The method firstly extracts the largest contour rectangle of every foreground pixel which holds the connected domain of the pixel. Then these contour rectangles are sorted by horizontal and vertical coordinates. Lastly, with the statistical information, the adjacent rectangles are merged. The whole handwritten Chinese manuscript is segmented to many domains. Every domain contains one or more Chinese character.This paper ultimately produces a handwritten character recognition system. The system implements every process from the acquisition of optical image, Chinese characters image preprocessing, the segmentation and standardization of Chinese characters, feature extraction and multi-classifiers ensemble. The system can recognize finally the Chinese characters of the handwritten Chinese manuscripts.