The Calligraphic Spectrum: Quantifying the Quality of Arabic Children’s Handwritten Character Generation Using CWGAN-GP and Multimeric Evaluation

Due to high intraclass variability and subtle intercharacter differences, automatic Arabic handwriting recognition remains a challenging task, particularly for children’s handwriting. This study proposes a hybrid framework that combines class-conditional Wasserstein generative adversarial networks with gradient penalty (CWGAN-GP) for data augmentation and a convolutional neural network (CNN) enhanced with squeeze-and-excitation (SE) blocks for improved feature discrimination. Experiments were restricted to disconnected (isolated) characters from the Hijja dataset, which comprised 12,355 samples divided as follows: 80% for training (9884), 10% for validation (1236), and 10% for testing (1235). Training the CNN on real data alone yielded an accuracy of 93.47%, while incorporating CWGAN-GP-generated samples improved performance to 96.27%. Notably, the proposed SE-CNN trained with the CWGAN-GP-augmented data achieved the highest accuracy of 99.27%. This result demonstrates that the combination of advanced generative data augmentation and architectural refinement significantly enhances Arabic handwritten character recognition performance.