Dynamic and Contextual Information in HMM Modeling for Handwritten Word Recognition

Area of Science:

• Computer Science
• Artificial Intelligence
• Pattern Recognition

Background:

• Handwriting recognition systems require robust modeling of writing units.
• Integrating dynamic and contextual information is crucial for improving accuracy.
• Existing methods may not fully leverage contextual dependencies in handwriting.

Purpose of the Study:

• To develop an efficient word recognition system by combining multiple handwriting recognizers.
• To enhance a Hidden Markov Model (HMM)-based recognizer with contextual information.
• To reduce model complexity while improving recognition performance.

Main Methods:

• A combined system using three handwriting recognizers, with a core HMM-based component.
• A state-tying process utilizing decision tree clustering for contextual unit modeling.
• Decision trees built using expert-based questions (global and precise) to cluster character models.

Main Results:

• Clustering reduced the number of models and Gaussian densities by a factor of 10.
• Contextual information, combined with dynamic modeling, significantly boosted recognition accuracy.
• Successful experiments conducted on diverse Latin and Arabic handwriting databases (Rimes, IAM, OpenHart).

Conclusions:

• The proposed HMM-based system effectively integrates dynamic and contextual information for superior handwriting recognition.
• Decision tree-based state-tying offers an efficient method for modeling contextual units in handwriting.
• The combined approach demonstrates significant improvements in handwritten word recognition across multiple languages.