Arabic calligraphy style recognition with edge enhancement and attention

Arabic calligraphy style recognition is a challenging task because visually similar styles often share curved strokes, connected structures, and subtle ornamental variations. This study proposes a convolutional neural network (CNN) architecture that combines learnable edge enhancement, squeeze-and excitation (SE) attention, and optional BiLSTM-based sequential modeling for Arabic calligraphy style classification. The edge module is designed to preserve fine-grained stroke boundaries, while the attention mechanism emphasizes informative channel responses and the BiLSTM models width wise contextual dependencies. Experiments were conducted on a public Arabic calligraphy style dataset using a stratified 70/15/15 train-validation test split. The best configuration achieved 98.22% accuracy and 98.20% F1-score on the test subset. These findings indicate that adaptive edge-aware feature learning can improve the discrimination of visually similar Arabic calligraphy styles. Nevertheless, broader validation on additional datasets and repeated runs is still needed to confirm generalizability.