TEYSuR - Text Extraction with YOLO and Super Resolution
Srividya Subramanian, Vineet Kekatpure, Gladina Raymond, Kapil Parab, S. Dugad, Archana Shirke · 2022 International Conference for Advancement in Technology (ICONAT) · 2022
In these times where a myriad of industries are becoming digitized and advanced, text recognition is being incorporated time and again to enhance customer satisfaction, improve accessibility and organize business processes. Conventional Deep Learning based text recognition, also termed as Optical Character Recognition (OCR) is based on learning the shapes of the characters in a language. In Text Extraction with YOLO and Super Resolution (TEYSuR) we propose a novel approach where YOLO, an object detection algorithm is used to localise and classify characters in an input image. The characters in the image are treated as individual objects by the algorithm and text is extracted with high accuracy and speed. In order to cater to input images of various font sizes the system proposes an additional Inference module which resizes images using Super Resolution and yields extremely high accuracy results. The word accuracy attained by the YOLO model was 88% and coupled with the Inference module TEYSuR attained a word accuracy of 96%. Hence, this system can be successfully used for extracting text from images with high efficiency.