Decoding the Devanagari - Handwritten Hindi Recognition Using Deep Learning Methodology

Recognizing text contained within images holds increasing significance as digital content incorporates more visual elements. The preservation and digital conversion of hand-written manuscripts are pivotal for their accessibility and integration with computing systems. Hindi, scripted in Devanagari, ranks as the world's third most prevalent language. However, research on optical character recognition for handwritten languages such as Hindi often lags behind that of English. This study attempts to address this disparity by presenting a method that employs Convolutional Neural Networks to identify handwritten Hindustani characters. The model, comprising eighteen layers, attained a 98.8% accuracy rate when trained on the Kaggle-provided Devanagari Character Set Classification dataset, encompassing over 92,000 images. Leveraging the Rectified Linear Unit (ReLU) and Leaky ReLU activation functions, the Adam optimizer, and categorical cross-entropy loss, the model effectively categorizes characters. This approach not only showcases its proficiency in accurately identifying handwritten Hindi characters but also ushers in new avenues for exploration and application within this domain. In future, advanced models like RNN, CNN, and LSTM can be harnessed to comprehend entire statements and beyond.