Authors: Alibaba Group (Hongsheng Jin, Mengli Cheng, Wengmeng Zhou, Qiangpeng Yang, Jun Huang, Wei Lin)
Description: A novel text spotter for arbitrary-shaped text detection and recognition is proposed by Alibaba-PAI OCR team. Specially, TextKeypointHead is introduced to detect more accurate text boundary with sixteen keypoints for text detection module. ROIThinPlateSpline is introduced to transform arbitrary-shaped features into rectified ones, then the rectified features can be recognized through a sequence recognition module. Benefiting from TextKeypointHead and ROIThinPlateSpline, our method is end-to-end trainable. The end-to-end training strategy learns more generic features to make our method perform better. In testing, multi-scale ensemble strategy is adopted.