ResTANet: Development of Handwritten Character Recognition using Adaptive Fused Neural Network for Ancient Tamil Document Analysis

Abstract The Tamil handwritten text is a significant in automatic document image analysis, despite prolonged research, Education and Multilingual Communication. In practical scenarios, identifying handwritten Tamil characters poses significant challenges due to the diverse writing styles adopted by individuals. The variations in writing make certain letters particularly challenging to decipher, and only a limited number of individuals may be proficient in comprehending these Tamil character. Here, The ResTANet is developed to improve the quality of the Tamil handwritten character recognition system. The dataset is collected via Mendeley website's (Tamil language dataset), and Gaussian Filtering with Otsu thresholding is used to filter and binarize the data to eliminate noise. We utilized and optimized robust deep learning methods including ResNet101 approach. The essential features are extracted using ResNet101. The parameters such as character length, width, and statistical features are optimized using Adam optimizer. Text quality was assessed experimentally using our model, and its performance is compared with other state of art algorithms and attains 99.37% accuracy. This type of analysis will be useful for document digitalization, Education, Historical document preservation and health care.