Minimum Error Rate Training for PHMM-Based Text Recognition

In this paper, discriminative training is studied to improve the performance of our Pseudo 2D Hidden Markov Models (PHMM)- based text recognition. In the traditional approach of model estimation, maximum likehood (ML) is usually used as the optimization criterion. Due to the lack of the correct information about the parametric forms of the observation models, ML-trained models do not necessarily achieve the minimum classification error (MCE). On the other hand, the aim of the discriminative training is to adjust model parameters to directly minimize the recognition error rate. PHMMs are optimized at work level, not only according to the correct token label, but also based on other most competitive word hypotheses. These competitive candidates are provided by the duration-corrected N-best hypotheses search, and the MCE optimization is implemented through the generalized probability descent algorithm. Experimental results have shown great reduction in recognition error rate through this training algorithm, and the performance is improved even for PHMM already been well-trained using conventional ML approaches