Bridging the Gap in Exam Handwritten Text Recognition: Dataset, Benchmark, and Modeling

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Handwritten text recognition (HTR) is crucial for intelligent grading systems.
• Existing HTR methods struggle with the unique challenges of exam handwriting, such as non-monotonic sequences and visual distortions.
• A systematic approach to benchmarking and robust HTR models for exam scenarios is needed.

Purpose of the Study:

• To systematically model and benchmark handwriting phenomena in examination settings.
• To develop a robust HTR model that overcomes the limitations of current methods in recognizing challenging exam handwriting.
• To provide a comprehensive evaluation framework for exam handwriting recognition.

Main Methods:

• Construction of BNU-Exam-HTR, a large-scale dataset of handwritten exam text.
• Establishment of BNU-Exam-Benchmark, an evaluation framework with 12 representative challenges.
• Proposal of EduOCR, a novel HTR model featuring a collaborative dual-branch decoder (Sequential Symbol Module and Permutation-Aware Prediction Head).

Main Results:

• EduOCR demonstrates superior performance compared to state-of-the-art HTR models, OCR tools, and multimodal large language models.
• The proposed model consistently outperforms existing methods across all 12 defined handwriting challenges.
• EduOCR exhibits enhanced robustness and adaptability in recognizing complex handwritten exam text.

Conclusions:

• EduOCR effectively addresses the challenges posed by exam handwriting artifacts, including sequential and visual disruptions.
• The developed dataset and benchmark provide a valuable resource for advancing HTR research in educational contexts.
• The proposed model offers a significant improvement for intelligent grading systems relying on accurate handwritten text recognition.