Devanagari Characters Recognition using Deep Neural Networks
Parveen Malik, Manoj Kumar Parida · 2022
Use of machines for human functions like reading, is always a dream for humans. Over the last few decades, machine reading has become a reality. Optical Character Recognition (OCR) is the most successful technology in this field. Handwritten or printed text recognition by the machine is termed as Optical Character Recognition (OCR). The demand for automatic character recognition systems has been increasing day by day and it can be seen in various forms like printed text or typed text as well as the handwritten one. This information can be further reprocessed for various applications and can be embedded with visual clues to enhance its usability in various applications. The optical character recognition is one such way, which can extract the information from printed documents. However, there are various technical challenges in accessing the data that is much worn out, old or specific to an object. A system trained on a particular model cannot inter-operate with other data, as there is no uniformity. The font size to writing style are specific to a person. Therefore, it is imperative to find a model that is universally accepted. In this paper, the Convolution Neural Network (CNN) as a machine-learning tool to classify the Devanagari characters is proposed. The automation provided by the deep learning models is utilized to eliminate the feature extraction step. The experimentation is done with various architectures of CNN with varying depth and structures and comparison has been done with various state of art methods like VGG16, VGG19, InceptionV3, MobileNet, ResNet50 and Xception etc. using transfer learning of pre-trained weights. The training set and validation set accuracy are found to be 99.90% and 99.09% respectively, which shows the prowess of proposed model.