Headline Recognition from Hindi Newspaper for Visually Impaired People using Deep Learning

Numerous technologies have been developed for the provision of visually impaired people. Amongst the several technologies, the solution provided by deep learning for computer vision is evolving as one of the most capable options owing to their affordability and availability. In this paper, we propose an end - to-end deep learning model to read the headlines of Hindi newspapers for visually impaired people. A headline is a significant element in a newspaper that provides the summary of the newspaper article. Moreover, there are about 186 million Indian readers of Hindi newspapers. The proposed system first extracts the headlines from the scanned newspaper articles using the You Only Look Once v4 (YOLOv4) object detector. The text from the scanned headlines is recognized using Convolutional Recurrent Neural Network (CRNN). Further, the extracted text is converted to sound that helps visually impaired people get important information from the Hindi newspapers. The results obtained have a WER of 1.8 % and a CER of 0.8 %. We have successfully converted all the text recognized by the system to sound to assist blind people in reading newspaper headlines.