Stroke Detector and Structure Based Models for Character Recognition: A Comparative Study.

Area of Science:

• Computer Vision
• Pattern Recognition
• Artificial Intelligence

Background:

• Characters are fundamental to communication, conveying semantic information through strokes and structure.
• Accurate character recognition is crucial for various applications, including scene text analysis and handwriting recognition.
• Existing part-based models and multi-scale stroke detectors often require manual labeling, limiting their scalability.

Purpose of the Study:

• To explore stroke and structure-based methods for enhanced character recognition.
• To develop automated methods for learning discriminative stroke detectors.
• To evaluate the performance of novel representations against existing models on challenging datasets.

Main Methods:

• Introduced two existing part-based models for character recognition.
• Proposed discriminative multi-scale stroke detector-based representation (DMSDR) to utilize multi-scale strokes.
• Developed discriminative spatiality embedded dictionary learning-based representation (DSEDR) for automatic learning of stroke detectors.

Main Results:

• Comparative study of Tree-Structured Model (TSM), mixtures-of-parts TSM, DMSDR, and DSEDR on scene character recognition (SCR) and handwritten digits datasets.
• Experimental results show stroke detector-based models are effective for characters with deformations and distortions.
• Demonstrated superior performance, particularly with limited training samples.

Conclusions:

• Stroke detector-based models offer a robust solution for character recognition challenges, including variations and limited data.
• Automated learning of stroke detectors via DSEDR overcomes limitations of manual labeling.
• The proposed DMSDR and DSEDR methods advance the state-of-the-art in character recognition systems.