Discriminative training methods and their applications to handwriting recognition

This thesis aims to improve the performance of handwriting recognition systems by introducing the use of discriminative training methods. Discriminative training methods use data from all competing classes when training the recogniser for each class. We develop discriminative training methods for two popular classifiers: Hidden Markov Models (HMMs) and a prototype-based classifier. At the expense of additional computations in the training process, discriminative training has demonstrated significant improvements in recognition accuracies from the classifiers that are not discriminatively optimised. Our studies focus on isolated character recognition problems with an emphasis on, but not limited to, off-line handwritten Thai characters. The thesis is organised as followed. First, we develop an HMM-based classifier that employs a Maximum Mutual Information (MMI) discriminative training criterion. HMMs have an increasing number of applications to character recognition in which they are usually trained by Maximum Likelihood (ML) using the Baum-Welch algorithm. However, ML training does not take into account the data of other competing categories, and thus is considered non-discriminative. By contrast, MMI provides an alternative training method with the aim of maximising the mutual information between the data and their correct categories. One of our studies highlights the efficiency of MMI training that improves the recognition results from ML training, despite being applied to a highly constrained system (tied-mixture density HMMs). Various aspects of MMI training are investigated, including its optimisation algorithms and a set of optimised parameters that yields maximum discriminabilities. Second, a system for Thai handwriting recognition based on HMMs and MMI training is introduced. In addition, novel feature extraction methods using block-based PCA and composite images are proposed and evaluated. A technique to improve generalisation of the MMI-trained systems and the use of N-best lists to efficiently compute the probabilities are described. By applying these techniques, the results from extensive experiments are compelling, showing up to 65% relative error reduction, compared to conventional ML training without the proposed features. The best results are comparable to those achieved by other high performance systems. Finally, we focus on the Prototype-Based Minimum Error Classifier (PBMEC), which uses a discriminative Minimum Classification Error (MCE) training method to generate the prototypes. MCE tries to minimise recognition errors during the training process using data from all classes. Several key findings are revealed, including the setting of smoothing parameters and a proposed clustering method that are more suitable for PBMEC than using the conventional methods. These studies reinforce the …