Confidence-Based Discriminative Training for Model Adaptation in Offline Arabic Handwriting Recognition

Abstract

We present a novel confidence-based discriminative training for model adaptation approach for an HMM based Arabic handwriting recognition system to handle different handwriting styles and their variations.Most current approaches are maximum-likelihood trained HMM systems and try to adapt their models to different writing styles using writer adaptive training, unsupervised clustering, or additional writer specific data.Discriminative training based on the Maximum Mutual Information criterion is used to train writer independent handwriting models. For model adaptation during decoding, an unsupervised confidence-based discriminative training on a word and frame level within a two-pass decoding process is proposed. Additionally, the training criterion is extended to incorporate a margin term.The proposed methods are evaluated on the IFN/ENIT Arabic handwriting database, where the proposed novel adaptation approach can decrease the word-error-rate by 33% relative.

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

“…It can be observed that both WER and CER are slightly decreasing with every Rprop iteration, and that between 10 and 15 Rprop iterations are optimal for the considered small and unsupervised labeled test datasets. Table 7 shows the final results of our Arabic handwriting recognition system with additional glyph dependent model length estimation (GDL) as described in [6]. Again, the WER of the GDL based system can be decreased by our proposed M-MMI training during both decoding passes down to 14.55%.…”

“…In addition to the novel confidence-based extension of the margin-based MMI training presented in [6], the confidence concept has been incorporated in the margin-based MPE criterion in this work. In the following, we give a brief summary.…”

“…In addition to the margin concept, the MMI/MPE training criteria are extended by an additional confidence term [6] to allow for novel unsupervised model adaptation.…”

Hierarchical hybrid MLP/HMM or rather MLP features for a discriminatively trained Gaussian HMM: A comparison for offline handwriting recognition

“…In [4,5] the authors apply a margin-based Maximum Mutual Information (M-MMI) and Minimum Phone Error (M-MPE) training criterion to an HMM-based off-line handwriting recognition system and report up to 33% relative improvement in word error rate (WER) compared to an ML trained baseline system.…”

Neural Networks for Document Image and Text Processing

“…Finally each image is re-sized to match the average pixel per character. For isolated words or scripts like Arabic, we could normalize each word to a fixed width keeping their aspect ratio [Dreuw et al 2009]. …”