MORAN: A Multi-Object Rectified Attention Network for scene text recognition

作者：

Highlights：

• We propose a new method, namely MORAN, to recognize irregular scene text.

• Trained under weak supervision manner, the sub-network MORN in MORAN is flexible. It is free of geometric constraints and can rectify images with complicated distortion.

• We propose a fractional pickup method to further improve the sensitivity of the attention-based decoder in MORAN.

• The proposed MORAN outperforms state-of-the-art methods on several standard text recognition benchmarks.

摘要

•We propose a new method, namely MORAN, to recognize irregular scene text.•Trained under weak supervision manner, the sub-network MORN in MORAN is flexible. It is free of geometric constraints and can rectify images with complicated distortion.•We propose a fractional pickup method to further improve the sensitivity of the attention-based decoder in MORAN.•The proposed MORAN outperforms state-of-the-art methods on several standard text recognition benchmarks.

论文关键词：Scene text recognition,Optical character recognition,Deep learning

论文评审过程：Received 11 March 2018, Revised 20 September 2018, Accepted 7 January 2019, Available online 15 January 2019, Version of Record 25 January 2019.

论文官网地址：https://doi.org/10.1016/j.patcog.2019.01.020