Off-line Handwritten Character Recognition Based On HMM

Chinese characters is one of the symbols of the Chinese civilization, with over one thousand years of history. Chinese character recognition has a wide range of application in office automation, machine translation, the news publishing and other areas. Off-line handwritten Chinese character recognition is the most difficult problem, and the identification process has several major steps including the preprocessing, classification, after handling, and so on. The target of preprocessing is to finish image binarization, denosing, character segmentation, thinning, and regularization. Feature extraction need to extract the feature that can effectively representing the character image. Classification is the key step in off-line handwritten Chinese character. The classifiers which are based on distance, neural network, and support vector machine (SVM) are the common classifiers. The main task of after handling is to make corrections using the context and improve the recognition rate further.In this paper, the preprocessing of the off-line handwritten Chinese character recognition is introduced, and the algorithms of traditional binarization, the smooth, character segmentation and refinement are discussed. The paper focuses on the handwritten character recognition classifier which is based on multiple hidden Markov model (MHMM), and analyses the structure of model and the methods and process of recognition classification. In addition, based on characters four directions projection, the paper picks up the boundary chain code features as the observation input of hidden Markov models, and uses the artificial neural network (ANN) for training, then proposes a handwritten recognition method which is based on artificial neural network multiple hidden Markov model (ANN-MHMM) combined the good resistance of neural network and the well-modeling advantages of hidden Markov model to improve the recognition performance.The recognition experiment has been performed by recognizing the money character of bank notes, the experiment results show that using the method this paper proposed of doing off-line handwritten character recognition is feasible, and the recognition accuracy rate is up to 96.3%, which is better than the existing methods.