End-to-End Scene Text Recognition with Character Centroid Prediction

Scene text recognition tries to extract text information from natural images, being widely applied in computer vision and intelligent information processing. In this paper, we propose a novel end-to-end approach to scene text recognition with a specially trained fully convolutional network for predicting the centroid and pixel cluster of each character. With the help of this new information, we can solve the character instance segmentation problem effectively and then combine the recognized characters into words to accomplish the text recognition task. It is demonstrated by the experimental results on ICDAR2013 dataset that our proposed method with character centroid prediction can get a promising result on scene text recognition.