COMPARATIVE ANALYSIS OF DIFFERENT MACHINE LEARNING MODELS ON HINDI OCR TASK

Abstract

The rapid advancement in Optical Character Recognition (OCR) technology offers promising solutions for digitizing printed and handwritten documents. However, OCR systems for Indian languages, particularly Hindi, have lagged due to the complexity of the Devanagari script. This project presents a comparative analysis of machine learning models, including Convolutional Neural Networks (CNN), XGBoost, and LightGBM, for the development of a robust Hindi OCR system. The study utilized a dataset of handwritten Hindi characters to train and evaluate the models. The CNN model, known for its proficiency in image classification tasks, achieved an accuracy of 99%, demonstrating its capability to learn and recognize complex patterns in Hindi script. LightGBM and XGBoost models, which were adapted to handle image data, attained accuracy of 80% and 84%, respectively. These results highlight the efficacy of gradient boosting algorithms in recognizing structured data, albeit with limitations in handling the intricacies of visual patterns compared to CNNs. The findings of this study underscore the potential of CNNs in developing accurate OCR systems for complex scripts like Devanagari, while also recognizing the competency of XGBoost and LightGBM in certain scenarios. This work contributes to the development of more inclusive OCR technologies, fostering greater accessibility to digital content in Hindi and other linguistically diverse languages. The project paves the way for future research to further enhance OCR accuracy and efficiency, exploring hybrid models that leverage the strengths of multiple machine learning techniques