Scene text recognition using similarity and a lexicon with sparse belief propagation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Natural Language Processing

Background:

• Scene text recognition (STR) faces challenges due to font variability, limited context, and uncontrolled environmental conditions.
• Existing methods often ignore valuable information like character image similarity or process data sequentially, leading to errors.
• Language priors can cause identical characters to be mislabeled.

Purpose of the Study:

• To develop a unified probabilistic model for STR that integrates character similarity, language properties, and lexical decision-making.
• To improve the accuracy and efficiency of STR by fusing multiple information sources within a single model.
• To eliminate unrecoverable errors caused by sequential processing in traditional STR systems.

Main Methods:

• Introduced a probabilistic model for STR that directly compares character images to ensure consistent labeling.
• Integrated language properties and lexicon information into the model for enhanced accuracy.
• Utilized sparse belief propagation for accelerated inference, reducing computational dependencies.

Main Results:

• Incorporating character similarity reduced character recognition error by 19%.
• Lexicon integration decreased word recognition error by 35%.
• Sparse belief propagation achieved a 12X speedup, reducing considered lexicon words by 99.9% with no accuracy loss.

Conclusions:

• Fusing diverse information sources within a single probabilistic model significantly enhances STR performance.
• The proposed model effectively addresses challenges posed by uncontrolled conditions and font variations in scene text.
• The method offers a more robust and efficient solution for real-world scene text recognition applications.