HMM-based lexicon-driven and lexicon-free word recognition for online handwritten Indic scripts

Area of Science:

• Computer Science
• Artificial Intelligence
• Natural Language Processing

Background:

• Online handwritten word recognition for Indic scripts lags behind Latin and Oriental scripts.
• Existing methods often lack script independence and data-driven approaches.

Purpose of the Study:

• To develop and evaluate novel, data-driven, and script-independent techniques for online handwritten word recognition in Devanagari and Tamil.
• To compare the performance of lexicon-driven and lexicon-free Hidden Markov Model (HMM) based approaches.

Main Methods:

• Proposed two HMM-based word recognition techniques: lexicon-driven and lexicon-free.
• Lexicon-driven technique: models words as sequences of symbol HMMs based on phonetic representation.
• Lexicon-free technique: utilizes a Bag-of-Symbols representation, independent of symbol order, for efficient lexicon pruning.

Main Results:

• A combination of lexicon-driven and lexicon-free recognizers significantly improved Devanagari recognition (87.13% accuracy for 20,000 words) on samples with varied writing orders.
• For Tamil, where standard symbol order predominates, the lexicon-driven recognizer achieved higher accuracy (91.8% for 20,000 words) than the lexicon-free approach or their combination.

Conclusions:

• Hybrid HMM approaches offer significant improvements for Devanagari handwritten word recognition, especially with diverse writing orders.
• The lexicon-driven HMM technique is highly effective for Tamil handwritten word recognition due to its consistent symbol order.