Authors: Yuliang Liu, Lele Xie, Lianwen Jin, Sheng Zhang

Description: We used a two stages deep learning method to roughly recall text first and then finely adjust the bounding box for compact detection. A novel network was carefully designed for multi-scale training, and we used a learnable method to generate quadrilateral proposals for tightly matching the scene text. Finally, we designed a post-processing pipeline for further improving the precision.