method: SCUT_DLVClab12017-06-28

Authors: Yuliang Liu, Lianwen Jin, Sheng Zhang, Canjie Luo, Zhaohai Li, Lele Xie, Zenghui Sun

Description: Two deep learning-based stages have been used, the first is to roughly detect text, and the second is to finely adjust the bounding box for accurate detection. A novel network is designed for multi-scale learning to generate tight quadrilateral proposals for the scene text. This is followed by a post-processing pipeline to improve the precision.