Authors: Xiang Chen (SOGOU), Yangchao Ou (SOGOU), Lulu Xu (SOGOU)

Description: The proposed method based on Mask-RCNN effectively detects text area by improving RPN and exploring the feature map of contextual information. To overcome the lack of the regression problem of targets with different scales, we improved the RPN based on GIOU and adopt an online hard example to better suppress false positives. At the same time, in order to obtain the characteristics of the context information, we add attention module.