Authors: Hailiang Xu, Feng Su, Tong Lu

Description: In text detection stage, seed candidate characters are detected with discriminative deep convolutional features learned within the maximally stable extremal regions extracted from the image, and are further grown to locate other degraded candidate characters. Then, the random forest classifier is exploited on the representative and distinct statistical features characterizing the properties of constituent character components in the text line, to predict the text and non-text label.
For text recognition, we exploit the CNN classifier to recognize the characters in each word. Then the levenshtein distance is used to find the nearest word in the dictionary in order to rectify some false character recognition.