method: Sensetime OCR2017-06-30

Authors: Xuebo Liu, Ding Liang, Dagui Chen, Minghao Guo, Junjie Yan

Description: A Deformable-FCN based network
Method description: An FCN-based method is used where for every pixel in a image, the method predicts whether it’s text, and also the text’s location if it’s in a text box. Online hard example mining and deformable convolution are used to improve the performance.