基于最佳鉴别变换的HMM手写数字字符识别

Handwritten Character Recognition Using HMM Based on Optimal Discriminant Transformation

基于隐马尔可夫模型 (HMM)的手写字符识别方法是近年来的一个研究热点 ,针对 HMM编码稳定性和建模过程复杂的问题 ,提出了一种新方法 ,即采用统计不相关最佳鉴别变换对模式进行特征抽取和降维 ,获得最佳鉴别特征向量 ,并在此基础上对各最佳鉴别方向的投影结果进行编码 ,作为 HMM的观测值序列 ,由于统计不相关最佳鉴别变换保证了变换特征向量集类内散布最小 ,类间散布最大的条件 ,使 HMM编码的稳定性和模式的可分性得到明显改善 ,通过对美国国家邮政局

Handwritten character recognition using the hidden Markov model (HMM) has been an active research topic for the past decade. One of the major problems, however, is that the handwritten characters may not exhibit consistent patterns due to different people's different writing styles. To enhance HMM's encoding stability and to reduce its modeling complexity, we propose a new approach in this paper. Specifically, we first obtain a set of uncorrelated optimal discriminant vectors by conducting feature extraction and dimension reduction using the uncorrelated Foley-Sammon transformation. Next, using a new feature space spanned by the optimal discriminant vectors, we obtain the projection coefficients of the raw data onto this new feature space. We then use these coefficients to form the observation sequence of the HMM. Because the uncorrelated Foley-Sammon transformation ensures minimum intra-class distance and maximum inter-class distance, it significantly improves HMM's encoding stability and difference classes' separability. In fact, the transformation allows different characters to be separable in many projection directions. To validate the accuracy and robustness of the proposed approach, we conduct experiments on the widely used US Postal Service (USPS) data set. Experiments show that the integration of the uncorrelated Foley-Sammon transformation and the HMM performs very well, achieving a recognition rate of 92%. It not only is better than regular HMM, but also is superior to the widely used nerual network based approaches.