End-to-End English Text Recognition In Natural Scene Image

Due to the rapid development of mobile Internet and the popularization of mobile electronic devices such as smartphone, people transmit and take the natural scene images conveniently. The text in the images contains a wealth of information, we hope that computer can detect and extract text information automatically without manual work. Unlike the traditional optical character recognition, text information extraction in natural scenes is challenged by the problem of varied font of text, different font layout, and complex background, etc. In recent years, research of scene text recognition has been significantly developed, but there is still a big gap between the widely application requirements and the achievements. So the research of end-to-end English text recognition method for natural scene is not only important to theoretical significance, but also has broad application prospects.The purpose of this paper is to develop some accurate and fast methods which can help to extract the accurate text position and correct text information from the natural scene images, and construct an end-to-end scene text recognition system. The extant methods have some weakness that either bet on few hand-crafted features or rely on heavy learning models, this paper carries out the research in three aspects:(1)The feature extraction of character based on unsupervised feature learning and hierarchical representation;(2) Text location in natural scene image;(3) The framework of an end-to-end text recognition system for natural scene. The main contributions of this paper are as follows:1. We propose the hierarchical representation of the character image features based on unsupervised feature learning. Firstly, we use the variant of K-means obtain the basis vectors from the training data, then extract character feature by combining convolutional neural network.2. We propose the framework of hierarchical text location by strings from characters to words. We adopt the Maximally Stable Extremal Region algorithm to extract character candidate regions. Next, obtaining character location by multilayer feature filtering method, then grouping the characters. In order to improve the accuracy of text detection, we design the structure feature of a string.3. We propose the algorithm framework about end-to-end scene text recognition and realize the experimental demonstration system. We integrate the text detection and character recognition modules, and design the text correction method based on the Edit Distance. The experimental results demonstrate the effectiveness of the proposed text recognition method.