关键词: 二阶隐马尔可夫模型, 笔画特征, 聚类, 数字识别

Abstract: An HMM based on blocks of original picture is proposed because of traditional Hidden Markov Model（HMM） being insufficient in modeling the structural information for pattern. This model uses each block of the original picture as the state, so it has a certain topology and can be used in modeling structural information of object. The second-order HMM with final-state is used in order to enhance the ability of description and the accuracy. The model is used in recognition of handwritten digits. Considering the structure of the handwritten digits and the topology of the model, a novel method of extracting the stroke feature is proposed, which extracts each feature vector of strokes with cross point. The recognition rate of 95.7% is obtained from the recognition of MNIST database.

Key words: second-order Hidden Markov Models, stroke feature, cluster, digits recognition