Multi-Attention Based Convolutional Neural Network for Tamil Handwritten Character Recognition

The robustness of a handwritten character recognition system hinges on its ability to generalize across diverse handwriting styles, especially in regional scripts like Tamil, which are known for their complex curves and structural intricacies. Despite Tamil’s cultural and linguistic significance, limited resources and datasets have posed challenges for developing effective recognition models. This paper presents a deep learning-based approach tailored specifically for recognizing Tamil characters using a custom-curated handwritten character dataset known as Handwritten Tamil Vowel-13(HTV-13). The dataset is collected from multiple individuals, incorporating variations in pen type, stroke intensity, lighting conditions, and writing orientations to simulate real-world scenarios. To address the challenges inherent in handwritten Tamil script recognition, an EfficientNet-B0 architecture enhanced with Squeeze-and-Excitation (SE) and Convolutional Block Attention Module (CBAM) was employed. These attention mechanisms enable the model to refine spatial and channel-wise features, resulting in a more discriminative representation of each character. The training process was further optimized using GPU acceleration and mixed-precision techniques to improve computational efficiency without sacrificing accuracy. This work marks a significant step toward developing a scalable, intelligent Tamil Optical Character Recognition (OCR) system. By leveraging attention-enhanced architectures and a representative dataset, the proposed framework lays the groundwork for future advancements in Tamil script digitization and serves as a promising direction for broader applications in Indic language technologies.