EfficientNetB2 for Recognizing Arabic Children's Handwriting: A Study on the ALPN Dataset

Automatic recognition of Arabic handwritten characters remains a challenging task due to the considerable variation in writing styles, particularly among children. This challenge is compounded when the system needs to process a broad range of handwritten elements, such as letters, numerals, and punctuation marks. Therefore, developing robust recognition systems capable of accurately classifying these elements is essential for supporting educational and digital applications. This study presents an Arabic children’s handwritten character recognition model based on transfer learning (TL) using the EfficientNetB2 deep learning (DL) architecture. The EfficientNetB2 was evaluated using two datasets: the publicly available Hijja dataset and the newly developed ALPN (Arabic Letters, Punctuation, and Numerals) dataset, which contains Arabic letters in different positional forms, Arabic and Hindi numerals, and punctuation marks collected from Libyan school children. A comprehensive pre-processing pipeline was applied, including automatic segmentation, Otsu binarization, line removal, contrast enhancement, normalization, and image resizing. The EfficientNetB2 model was trained and evaluated using a hold-out validation strategy and further validated using 10-fold cross-validation (CV). Experimental results demonstrated that the built model achieved an accuracy of 88.25% on the Hijja dataset and 95.67% on the ALPN dataset using the hold-out validation method. Furthermore, using the cross-validation (CV) method, the findings confirmed the model’s robustness, achieving 88.81% and 95.58% accuracy on Hijja and ALPN, respectively. The obtained results indicate that EfficientNetB2 is highly effective for Arabic children’s handwriting recognition and that the ALPN dataset provides a reliable benchmark for future research in this field.