Combining Structure and Parameter Adaptation of HMMs for Printed Text Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Hidden Markov Models (HMMs) are widely used for sequence modeling.
• Existing HMM adaptation methods like Maximum A Posteriori (MAP) and Maximum Likelihood Linear Regression (MLLR) focus on parameter tuning.
• Adapting HMM structure alongside parameters can potentially improve model performance.

Purpose of the Study:

• To introduce two novel semi-supervised algorithms that extend existing HMM adaptation techniques.
• To integrate HMM structure optimization with parameter adaptation (MAP and MLLR).
• To improve the accuracy of printed character recognition by adapting HMMs to new fonts.

Main Methods:

• Developed semi-supervised algorithms combining MAP/MLLR with HMM structure optimization.
• Employed state splitting and merging operations for structure optimization based on likelihood or heuristic criteria.
• Utilized limited labeled data for parameter adaptation and moderate unlabeled data for structure optimization criteria estimation.
• Applied algorithms to adapt HMM character models for a polyfont printed text recognizer.

Main Results:

• Achieved significant increases in printed character recognition accuracy compared to standard MAP and MLLR.
• Successfully adapted HMM character models to new fonts using the proposed methods.
• Experiments involved a large dataset of 1,120,000 real and 3,100,000 synthetic character images across 89 HMM models.

Conclusions:

• The proposed algorithms effectively enhance HMM adaptation by incorporating structure optimization.
• Semi-supervised learning enables efficient adaptation with limited labeled and moderate unlabeled data.
• The approach offers a significant improvement in character recognition accuracy, particularly for adapting to new fonts.