TifinNet: CNN–Transformer Hybrid Architecture for Tifinagh Handwritten Recognition

Handwritten Optical Character Recognition (OCR) for the Tifinagh script remains a challenging task due to the script’s geometric complexity, high similarity among rotated characters, and the scarcity of annotated training data. Existing methods often struggle to handle the directional features and orientation variations characteristic of Tifinagh handwriting. To address these challenges, we propose TifinNet, a novel CNN–Transformer hybrid architecture tailored for handwritten Tifinagh character recognition. The model integrates geometry-aware convolutional layers with multi-scale attention mechanisms, enabling it to capture both fine-grained stroke details and broader contextual dependencies. Key contributions include the introduction of oriented convolution operations for modeling directional features across multiple angles, spatial attention modules to highlight character-relevant regions, and a hybrid training scheme that combines CTC and attention-based losses. Experimental results on the AMHCD dataset demonstrate that TifinNet achieves state-of-the-art accuracy while maintaining computational efficiency. An ablation study further validates the importance of the oriented convolution layers and the hybrid loss design. At the sequence level, the model delivers higher word recognition accuracy and lower edit distances compared to existing approaches, highlighting its practical effectiveness for end-to-end text recognition. Overall, TifinNet provides a robust and efficient solution for Tifinagh OCR and offers broader insights into the design of CNN–Transformer architectures for low-resource script recognition.