method: A Text Localization Method Based on CTPN2019-04-20

Authors: Songyi Yang, Shengjie Xiu, Niansong Zhang

Description: A novel Connectionist Text Proposal Network (CTPN) published by Tian, Zhi, et al. is deployed in this model. The CTPN model develops a vertical anchor mechanism for character detection whose result is refined to generate an accurate bounding box prediction.

Most state-of-art algorithms are based on the bounding box regression or character-level feature extraction. However, this method combines the character-level bounding box regression with label prediction. A CNN is naturally used as a feature extractor, whose output is divided and fed into a bi-directional LSTM and fully -connected layer for anchor prediction and label classification. The end-to-end trainable model works with the various scale of texts even with some extremely ambiguous text.

It achieves 0.88 and 0.61 F-measure on ICDAR 2013 and 2015 benchmarks, according to the original paper. The publication and our implementation is given in the following sections.