Handwriting Digital Recognition via Modified Logistic Regression

Motivated by a wide range of real world applications of hand writing digital recognition, e.g., postal code recognition, the past decades have seen its great progress. The related approaches are generally composed of two components, feature extraction and identification methods. We note that the previous approaches are limited by the following two aspects: (1) the feature is not adaptive enough to cover the great variance within data; (2) the recognition methods are suffered from local minima solution. Inspired by these observations and to overcome these limitations, we in this paper propose an approach HMM-MLR by exploiting hidden Markov model (HMM) and modified logistic regression (MLR). In the proposed approach, HMM is employed to model the trace of handwriting digital, which is able to model the large variance within digitals and can adapt to the data distribution. Then the features are extracted based on HMM and then delivered into MLR for recognition. Benefitting from the global optimum solution of MLR, the proposed approach could reach highly stable results. To verify the effectiveness of the proposed approach, we experimentally compare our proposed approach with other state-of-the-art approaches over Semeion handwritten digit dataset. The experimental results show that, over both recognition accuracy and recall, for different rounds of experiments and different number of training samples, our HMM-MLR exhibits significant improvement over others.