An Attention-Driven CNN-GRU Framework for Handwritten Character Recognition

Optical Character Recognition (OCR) is a process that is used to convert the handwritten text into a digital format so that computers can understand. Telugu, one of the widely used languages in South India, has more complex and similar characters, which bring challenges to the OCR systems. To overcome these challenges, this study introduced a hybrid deep learning approach that combines Convolution Neural Network (CNNs) for the extraction of the pictorial features, Bidirectional Gated Recurrent Units (Bi-GRU) for understanding sequential information and an additive attention Mechanism to high-light important features. This hybrid model is trained and tested on well-known datasets and achieved an accuracy of 97.396% with a test loss of 0.1050. This shows how this method makes OCR systems more precise and reliable when working with challenging scripts such as Telugu.