[Paper Review] Handwritten Arabic Character Recognition for Children Writ-ing Using Convolutional Neural Network and Stroke Identification

Mais Alheraki, Rawan Al-Matham+1 more|arXiv (Cornell University)|Nov 3, 2022

Handwritten Text Recognition Techniques4 citations

TL;DR

This paper proposes a convolutional neural network (CNN) model enhanced with stroke-based classification for recognizing handwritten Arabic characters written by children. It achieves 91% accuracy on the Hijja dataset and 97% on the Arabic Handwritten Character Dataset (AHCD), with a multi-model approach using stroke count improving average accuracy to 96% when combining both datasets.

ABSTRACT

Automatic Arabic handwritten recognition is one of the recently studied problems in the field of Machine Learning. Unlike Latin languages, Arabic is a Semitic language that forms a harder challenge, especially with variability of patterns caused by factors such as writer age. Most of the studies focused on adults, with only one recent study on children. Moreover, much of the recent Machine Learning methods focused on using Convolutional Neural Networks, a powerful class of neural networks that can extract complex features from images. In this paper we propose a convolutional neural network (CNN) model that recognizes children handwriting with an accuracy of 91% on the Hijja dataset, a recent dataset built by collecting images of the Arabic characters written by children, and 97% on Arabic Handwritten Character Dataset. The results showed a good improvement over the proposed model from the Hijja dataset authors, yet it reveals a bigger challenge to solve for children Arabic handwritten character recognition. Moreover, we proposed a new approach using multi models instead of single model based on the number of strokes in a character, and merged Hijja with AHCD which reached an averaged prediction accuracy of 96%.

Motivation & Objective

• Address the challenge of recognizing handwritten Arabic characters written by children, a less-studied demographic in Arabic HTR.
• Improve recognition accuracy for children's handwriting, which exhibits high variability due to age-related writing patterns.
• Develop a multi-model approach based on stroke count to better capture structural differences in children's handwriting.
• Enhance performance by combining the Hijja and AHCD datasets to increase training diversity and robustness.
• Demonstrate the effectiveness of CNNs with stroke identification for low-resource, age-specific handwritten Arabic character recognition.

Proposed method

• Propose a single CNN model trained on the Hijja and AHCD datasets to recognize Arabic characters written by children.
• Implement a stroke-based classification strategy, dividing characters into sub-models based on the number of strokes in each character.
• Fine-tune the CNN architecture to extract hierarchical spatial features from handwritten character images.
• Train and validate the model using data augmentation and transfer learning techniques to improve generalization.
• Combine predictions from multiple stroke-based models to improve overall recognition accuracy.
• Merge the Hijja and AHCD datasets to increase training data diversity and improve model generalization across writing styles.

Experimental results

Research questions

• RQ1 How does a CNN-based model perform in recognizing Arabic handwritten characters written by children compared to existing methods?
• RQ2 To what extent does incorporating stroke count as a feature improve recognition accuracy for children's handwriting?
• RQ3 Can combining the Hijja and AHCD datasets lead to better generalization and higher accuracy in children's Arabic HTR?
• RQ4 How does the multi-model approach based on stroke count compare to a single unified model in terms of performance and robustness?
• RQ5 What are the key challenges in recognizing children's Arabic handwriting, and how can they be mitigated using deep learning and structural analysis?

Key findings

• The proposed CNN model achieved 91% accuracy on the Hijja dataset, significantly outperforming the baseline model from the dataset authors.
• On the Arabic Handwritten Character Dataset (AHCD), the model reached 97% accuracy, demonstrating strong performance on a broader handwriting distribution.
• The multi-model approach, which splits characters by stroke count and uses separate models, improved average prediction accuracy to 96% when combining Hijja and AHCD.
• The results indicate that stroke-based modeling enhances recognition by capturing structural variations unique to children's handwriting.
• Despite improvements, the study identifies ongoing challenges in recognizing children's Arabic handwriting due to high variability in writing patterns.
• The combination of Hijja and AHCD datasets significantly improved model generalization, suggesting data diversity is critical for robust HTR in low-resource settings.