ODIA HANDWRITTEN OPTICAL CHARACTER RECOGNITION USING TRANSFER LEARNING AND PRE-TRAINED TESSERACT ODIA DATASET

Abstract

This study aims to address the problem of Optical Character Recognition (OCR) for the Odia language using a transfer learning.OCR is a technology used to convert different types of documents, such as scanned paper documents, PDF files or images captured by a digital camera, into editable and searchable data. The Odia language, like many other languages, has unique characteristics and complexities in its script that pose challenges for OCR. Transfer learning, which involves applying knowledge learned from one problem to a different but related problem, is seen as a potential solution to these challenges. This technique is especially beneficial when the dataset for the specific task (in this case, Odia OCR) is small, as it leverages the knowledge captured by models pre-trained on larger, more diverse datasets. In this study, a pre trained Convolutional Neural Network (CNN) model for feature extraction. CNNs are a type of deep learning model that are particularly good at processing grid-like data, such as images. A pre-trained CNN model is a model that has been previously trained on a large dataset, usually on a general task like identifying objects within images. The learned weights of this model, which capture the learned features from the previous task, are then used as the starting point for the new task. After initializing the model with the pre-trained weights, the authors fine-tune it on the specific task of recognizing Odia characters. Fine-tuning involves continuing the training process on the new task, adjusting the weights of the model to better fit the new data. The specifics of fine tuning, such as which layers of the model to fine-tune and the learning rate to use, can vary depending on the specifics of the task and the amount of available data. The dataset used in this study consists of images of 8 unique vowels in the Odia language. Image datasets for deep learning often require preprocessing to ensure that they can be efficiently and effectively fed into the model. In this case, the authors applied several preprocessing techniques like Image re-sizing, normalization, data splitting. The study thus offers a comprehensive approach to tackling the problem of Odia OCR using transfer learning, from using a pre-trained CNN model to fine-tuning it on a specific dataset, and meticulously preparing the data for optimal results.