A New Handwritten Number Recognition Method Using HMM Based on MNIST

Taking MNIST data-set as research object, the HMM is introduced into the Handwritten Number Recognition for the first time. After the implementation of the classical HMM training algorithm, some optimization methods are proposed for the problems existing in the training. The general random initialization parameters lead to long training time and unstable data. The training of initialization parameters based on observations can speed up the training and avoid data overflow. The number of iterations in the training process is not positively related to the output probability. In order to obtain an optimal model, after the algorithm converges when the cross entropy loss function of the output probability of two adjacent iterations is the minimum, the training ends. Finally, a comparative experiment between the classical method and the optimization method is carried out in the two stages of training and recognition. In the training stage, compared with the classical method, the optimization method has the advantages of short average training time, high average output probability, fast iteration speed and high accuracy. In the test stage, the accuracy of the optimization method is higher than that of the classical method. The experimental results show that the HMM can be effectively applied in the field of Handwritten Number Recognition. And through the optimization method, to a certain extent, the recognition accuracy is improved. In a word, the method in this paper is an effective and feasible method.