Image-Based Recognition of Children's Handwritten Arabic Characters Using a Confidence-Weighted Stacking Ensemble

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Handwritten Arabic character recognition, especially for children, is complex due to writing variations and diacritics.
• Deep learning models show promise but struggle with reliability in this domain.

Purpose of the Study:

• To develop a robust stacking ensemble framework for sensor-acquired Arabic handwriting data.
• To enhance prediction reliability using a dynamic confidence-thresholding mechanism.

Main Methods:

• Integrated three Convolutional Neural Networks (ConvNeXtBase, DenseNet201, VGG16) using a fully connected meta-learner.
• Implemented an optimized confidence threshold to filter uncertain predictions, maximizing the F1 score.
• Evaluated the framework on the Hijja and Dhad benchmark datasets for children's Arabic handwriting.

Main Results:

• Achieved state-of-the-art performance: 95.13% accuracy and 94.62% F1 score on Hijja.
• Achieved 96.14% accuracy and 95.59% F1 score on Dhad.
• Demonstrated over 3% accuracy improvement on Hijja compared to existing methods.

Conclusions:

• Confidence-based stacking ensembles effectively enhance reliability in Arabic handwriting recognition.
• The proposed framework shows strong potential for automated educational assessment and intelligent tutoring systems.