Arabic Handwriting Recognition and Writer Identification Using a CRNN-Based Deep Learning Framework on the KHATT Dataset

الملخص

Abstract—This study proposes a complete system for Arabic handwriting recognition and biometric writer identification using deep learning. Arabic handwriting faces challenges due to its cursive nature, position-dependent character shapes, and diacritical marks. A Convolutional Recurrent Neural Network (CRNN) architecture combing CNN, BiLSTM and CTC was designed and trained on the KHATT dataset. A label quality issue was identified during preprocessing and corrected, resulting in a 10-percentage-point improvement in Character Error Rate (CER), reducing it from 91.66% to 81.38% after 80 training epochs. An SVM classifier was incorporated for biometric writer identification based on CNN-extracted features, achieving a Top-5 accuracy of 78.85% across 20 writers.