Writer Identification of Arabic Historical Document Using a Deep Learning Approaches

Source abstract

Historical documents contain great information for scientific and literary research. Many documents suffer from degradation, especially on initial pages, making identifica- tion difficult when no attribution exists.Arabic historical documents have two challenges: Complexity of the script and poor physical condition. We address the problem of identity loss in Arabic historical documents by presenting a deep learning-based approach. We used a subset of the WAHD dataset comprising 16,491 images: known authors 60% and unknown authors 40%.Data augmentation was applied to enhance diversity. The data was split into 70% for training, 10% testing, and 20% validation. We implemented two models:The first, Deep Writer, is a deep convolutional neural network with a dual-path architecture, consisting of multiple convolutional, pooling, and fully con- nected layers. The second, Half Deep Writer, a similar structure but uses a single pipeline. We experimented different learning rates and found 0.0001 and 0.0002 gave optimal results. Model performance was evaluated using precision, recall, and F1-score to handle class imbalance. The Deep Writer model achieved 92.28% accuracy and an F1-score of 81.16%, while the Half Deep Writer model achieved 92.10% accuracy and an F1-score of 81.63% at a learning rate of 0.0002.