ResTANet: A Deep Residual Neural Architecture for Tamil Handwritten Character Recognition

In education, multilingual communication, and the preservation of historical documents, handwritten Tamil character recognition plays a crucial role in automatic document image analysis. Although extensive research has been conducted on handwritten Tamil text recognition, existing methods are still limited by the high variability in individual writing styles and the inherent complexity of the script. As a result, traditional deep learning approaches are often insufficiently robust to handle noise and variations. To address these challenges, we propose ResTANet, an enhanced deep learning framework for Tamil handwritten character recognition. In our approach, we use a modified version of ResNet101 to extract features, incorporating architectural optimizations to enhance generalization across different styles of handwriting. This study employed our own dataset sourced from Mendeley, https://data.mendeley.com/datasets/cpr8r8m2v9/1, encompassing a wide range of writing styles. The pre-processing steps included Gaussian filtering followed by Sauvola’s adaptive binarization procedure, which proved effective for handling uneven illumination and complex background noise. Feature extraction focused on geometric and statistical descriptors, which are subsequently optimized using the Adam optimizer. Based on experimental evaluation, the proposed ResTANet achieves an accuracy of 99.37%, a precision of 92.71%, a recall of 92.61%, and an F1 score of 92.71%. This model performs better than state-of-the-art models in terms of accuracy, precision, recall, and F1-score. This study provides implications for document digitization, educational resource development, historical manuscript preservation, and healthcare informatics applications.