DEEP LEARNING BASED FULL END-TO-END TEXT RECOGNITION SYSTEM

Abstract

With everything being automated in recent years, fields like speech recognition, object recognition and text recognition have attracted attention of majority of researchers in field of machine learning. Text detection and recognition specially have received greater attention due to “robust reading competitions” held over the years. End-to-end text recognition specially on natural scene images is a difficult task, as compared to that on documents, that has received attention of researchers in recent times. Traditional approaches in this area have been proposed with hand-engineered features and supervised or semi-supervised learning. In this project, I have combined a powerful 2-layered neural network with recent advances in the field of unsupervised machine learning, which is highly responsible for development of a module consisting of an accurate text detector followed by a character recognizer. Then using post processing methods like NMS are responsible for a full end-to-end text detector and character recognizer system, giving close to state-of-the-art performance on ICDAR and SVT datasets