UltraTamNet lightweight tamil-optimized deep learning architecture for handwritten character recognition

Tamil language, derived from the Brahmi script, is one of India’s oldest languages, Malaysia, Sri Lanka, Singapore, etc., with a history spanning over 3000 years. The Tamil script is a very complex script consisting of 247 characters with various shapes, including lines, curves, loops, and similarities in appearance between several characters, making it difficult to classify. Image processing methods were used before the use of Deep Learning (DL) and Machine Learning (ML) algorithms. These methods were lightweight but insufficiently reliable to classify handwritten Tamil characters in a variety of settings, such as lighting and orientation. A few ML or DL based algorithms were used to classify handwritten Tamil characters, but with a tradeoff between lightweight and accuracy. To address these limitations, this research work introduces UltraTamNet. This Tamil-optimised hybrid CNN model combines depthwise separable and residual blocks to capture complex characters with minimal computational requirements, thereby achieving high accuracy. The proposed method extracts significant information from handwritten Tamil characters by employing a series of effective feature-extraction layers. The UltraTamNet model is tested on two datasets, with a public uTHCD, and a customized dataset that was gathered from native writers. The UltraTamNet model outperforms current SOTA models. According to experimental results, on average, the proposed model achieved a 1.3% higher accuracy than the average accuracy of all compared architectures. It achieves 99.5% accuracy on the training set and 98.2% accuracy while keeping a lightweight 1.27 M parameter footprint. The findings demonstrate that the UltraTamNet model strikes a balance between accuracy and computational efficiency to enhance performance in Tamil handwritten character recognition. All source code required to reproduce the UltraTamNet model architecture and evaluation scripts are available in 1 .