Authors: Chen Du

Description: We develop a novel method for robustly recognizing both regular and irregular natural texts. The proposed irregular text recognition method consists of three stages: text direction classification, irregular text rectification, and sequence recognition. Text direction classification aims to judge the text orientation. The Irregular-text-rectification stage employs a centerline rectify algorithm to make the curved text horizontal. Then the rectified text image is recognized by an attention-based sequence recognition network.