method: StradVision2015-03-23

Authors: Hojin Cho et al.

Description: First, we extract character candidates using extremal regions (ER) Second, we verify the extracted character candidates with the character classifier trained by Agile Learning. Afterwards, we do text-patch matching which greatly enhances the recall rate, and group the characters into text regions. Finally, we apply a deep neural network for character recognition.

method: BUCT_YST2015-01-12

Authors: Wei Hu

Description: Multiple MSERs and Convolution Neural Networks are employed in our method to extract character candidates. Character candidates are then grouped by using the technique in USTB_TexStar to generate text candidates. At last, a text classifier is trained on the training dataset to verify these text candidates. The related demo of the software has been released on ( or

method: I2R_NUS_FAR2013-04-09

Authors: Lu Shijian, Tian Shangxuan, Lim Joo Hwee, Tan Chew Lim

Description: This method builds upon I2R_NUS by further reducing false alarms through a machine learning approach.

Ranking Table

Description Paper Source Code
DateMethodPx. RecallPx. PrecisionPx. F-scoreWell s.MergedBrokenBr.-Mer.LostFalse p.DetectedRecallPrecisionFscore
2013-04-06Text Detection64.74%76.20%70.01%372433925417761885633863.46%57.97%60.59%

Ranking Graphic

Ranking Graphic