Application Research On BWS-SOM Model In Chinese Recognition In Large Character Set

Chinese characters,which are used most in the world,are more difficult than others languages for recognizing due to its large character set,high similarity,font diversity and so on.With the development of pattern recognition,Chinese character recognition technology has grown up as an important branch of pattern recognition.After years of research,recognition technology for small character set has been developed,but for large character set,there are still some problems needed to improve.At present,in the application area of machine vision,Optical Character Recognition(OCR)has been a mainstream technology for large character set recognition,but its accuracy,for Chinese characters with similar structure or unclear Chinese characters,can be easily affected by picture quality,scanning software and scanning process.With the development of artificial intelligence,neural network recognition method have become one of the most important method of Chinese character recognition,but this method may always be influenced by the numbers of Chinese character so as to cause large scale network,most researches have been focusing on small character set.Self-organizing Maps(SOM)neural network,which can simulate self-organization characteristics of the brains while processing information,has been widely used in pattern recognition as its simple structure,the characteristics of self clustering and dimension reduction.While traditional SOM neural network is used in pattern recognition,the number of neurons in the input layer is equal to the number of pixels of sample picture,and the number of neurons in the output layer is equal to or greater than the number of categories.Therefore when it is used in large character set recognition,such as Chinese characters,the scale of the SOM neural network will get very large.In order to solve the above problems,a Blocked Winner Sequence based on SOM(BWS-SOM)model is proposed for Chinese recognition in large character set.Compared with others recognition methods based neural network,BES-SOM model improves as follows:(1)Image block and SOM neural network are combined for extracting the minutiae features of the image and reducing the number of neurons in the input layer,and then reducing the scale of network;(2)The ordered combination of multiple winner neurons are used to characterize the feature,which can greatly reduce the scale of model and improve classification ability of the model.To demonstrate the feasibility of BWS-SOM model,it is applied to the characters from Level 1 national standard Chinese character library.Firstly the separability of features extracted by the model and classification ability of the model are verified;Secondly the BWS-SOM model is applied to single font Chinese recognition in large character set;Finally based the best pre-treatment method and block method,the BWS-SOM model is applied to multi-font Chinese recognition in large character set.Experimental results show that while dealing with the same classification problem,BWS-SOM model has small network scale,less computation and higher recognition rate than traditional SOM network and other networks.It is effective in Chinese character recognition with rate 98.05% on single font with 97 neurons and 89.12% on multi-font with 59 neurons.