Detection and rectification of arbitrary shaped scene texts by using text keypoints and links

作者：

Highlights：

• We propose a robust scene text detection and rectification technique that is capable of detecting and rectifying scene texts of arbitrary shapes almost simultaneously.

• We formulate scene text detection and rectification as a text keypoint and link detection problem and proposes a mask-guided multi-task network that is capable of detecting text keypoints and keypoint links accurately.

• We develop an efficient and end-to-end trainable system that achieves superior scene text detection and rectification performance as compared with the state-of-the-art.

摘要

•We propose a robust scene text detection and rectification technique that is capable of detecting and rectifying scene texts of arbitrary shapes almost simultaneously.•We formulate scene text detection and rectification as a text keypoint and link detection problem and proposes a mask-guided multi-task network that is capable of detecting text keypoints and keypoint links accurately.•We develop an efficient and end-to-end trainable system that achieves superior scene text detection and rectification performance as compared with the state-of-the-art.

论文关键词：Scene text detection,Scene text recognition,Deep learning,Neural network

论文评审过程：Received 26 November 2020, Revised 2 June 2021, Accepted 4 December 2021, Available online 20 December 2021, Version of Record 23 December 2021.

论文官网地址：https://doi.org/10.1016/j.patcog.2021.108494