Maximization of mutual information for offline Thai handwriting recognition

Abstract

This paper aims to improve the performance of an HMM-based offline Thai handwriting recognition system through discriminative training and the use of fine-tuned feature extraction methods. The discriminative training is implemented by maximizing the mutual information between the data and their classes. The feature extraction is based on our proposed block-based PCA and composite images, shown to be better at discriminating Thai confusable characters. We demonstrate significant improvements in recognition accuracies compared to the classifiers that are not discriminatively optimized.

Confidence- and margin-based MMI/MPE discriminative training for off-line handwriting recognition

“…, T which are augmented by their spatial derivatives in horizontal direction ∆ = x t − x t−1 . Note that many systems divide the sliding window itself into several sub-windows and extract different features within each of the sub-windows [2,20,30,39] In order to incorporate temporal and spatial context into the features, we concatenate 7 consecutive features in a sliding window with maximum overlap, which are later reduced by a PCA transformation matrix to a feature vector x t of dimension 30 (see Figure 2). …”

“…In general, the number of Rprop iterations and the choice of the regularization constant C have to be chosen carefully (cf. optimization Table 1 in Section 2.3), and were empirically optimized in informal experiments to 30 Rprop iterations and C = 1.0 (cf. detailed Rprop iteration analysis and convergence without overtraining in Figure 8 and Figure 9).…”

“…The MMI training criterion has been used in [30] to improve the performance of an HMM based offline Thai handwriting recognition system for isolated characters. They propose a feature extraction based on a blockbased PCA and composite image features, which are reported to better at discriminating Thai confusable characters.…”

Markov models for offline handwriting recognition: a survey

“…Although certain publications regarding Markov-modelbased recognition of isolated characters exist (cf., e.g., [21,56,57,95]), it is at least questionable whether the use of these models is appropriate for such data. Instead, the approach shows its strength especially for sequences.…”

“…Often (local) pixel intensities or pixel density distributions-optionally their average or median values-are considered as some sort of basic features (as, e.g., in [3,40,45,87,117]). Quite frequently, starting from some raw feature set optimized representations are computed by some standard analytic transforms like PCA [28,40,58,95,97,117] LDA [16,58] or function transforms (DCT, FFT, Wavelet, Radon, etc.) [16,30,45].…”