Learning Spatial-Semantic Context with Fully Convolutional Recurrent Network for Online Handwritten Chinese Text...

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Online handwritten Chinese text recognition (OHCTR) presents significant challenges due to large character sets, ambiguous segmentation, and variable sequence lengths.
• Existing methods often struggle with the inherent complexities of handwritten input, limiting recognition accuracy.

Purpose of the Study:

• To develop a robust and accurate system for OHCTR.
• To overcome the limitations of traditional segmentation-based approaches.
• To leverage advanced deep learning techniques for improved performance.

Main Methods:

• Utilizing path signatures to transform pen-tip trajectories into informative feature maps, capturing stroke properties.
• Proposing a multi-spatial-context fully convolutional recurrent network (MC-FCRN) to process these feature maps and generate predictions without explicit segmentation.
• Developing an implicit language model to incorporate semantic context and lexicon constraints for enhanced prediction.

Main Results:

• Achieved high accuracy rates of 97.50% on Dataset-CASIA and 96.58% on Dataset-ICDAR.
• Demonstrated superior performance compared to existing state-of-the-art methods in OHCTR.
• Successfully addressed the segmentation problem through the proposed network architecture.

Conclusions:

• The proposed path signature and MC-FCRN approach offers a significant advancement in OHCTR.
• The integration of an implicit language model effectively utilizes linguistic knowledge for improved recognition.
• This method provides a robust and accurate solution for real-world applications of handwritten Chinese text recognition.