rflearning

RF-Learning

1. Introduction

This code repository contains an implementation of (RF-Learning: Reciprocal Feature Learning via Explicit and Implicit Tasks in Scene Text Recognition (ICDAR 2021)) . We excavate the implicit task, character counting within the traditional text recognition, without additional labor annotation cost. The implicit task plays as an auxiliary branch for complementing the sequential recognition. We design a two-branch reciprocal feature learning framework in order to adequately utilize the features from both the tasks. Through exploiting the complementary effect between explicit and implicit tasks, the feature is reliably enhanced.

2. Preparing Dataset

Train Dataset

Dataset	Samples	Description	Release
MJSynth	8919257	Scene text recognition synthetic data set	Link
SynText	7266164	A synthesized by scene text dataset, and the text is cropped from the large image	Link

Validation Dataset

Test Set	Instance Number	Note
IIIT5K	3000	regular
SVT	647	regular
IC03_860	860	regular
IC13_857	857	regular
IC15_1811	1811	irregular
SVTP	645	irregular
CUTE80	288	irregular

Test Dataset

Test Set	Instance Number	Note
IIIT5K	3000	regular
SVT	647	regular
IC03_860	860	regular
IC13_857	857	regular
IC15_1811	1811	irregular
SVTP	645	irregular
CUTE80	288	irregular

3. Getting Started

Preparation

A quick start is to use above lmdb-formatted datasets that contain the full benchmarks for scene text recognition tasks as belows.

Data Type: LMDB

File storage format:
   |-- train           
   |   |-- MJ
   |   |-- ST
   |-- validation
   |   |-- mixture
   |-- evaluation
   |   |-- mixture

Training

Run the following bash command in the command line,

cd .
bash ./train.sh 

We provide the implementation of online validation. If you want to close it to save training time, you may modify the startup script to add --no-validate command.

Evaluation

1.Visual Stage

  cd .
  bash ./test_scripts/test_rfl_visual.sh

2.Total Stage

  cd .
  bash ./test_scripts/test_rfl_total.sh

4. Results

Evaluation

Methods	Regular Text				Irregular Text			Download
Name	IIIT5K	SVT	IC03	IC13	IC15	SVTP	CUTE80	Config	Model
RF-Learning visual(Report)	95.7	94.0	96.0	95.2	84.2	87.0	85.8	-	-
RF-Learning visual	96.0	94.7	96.2	95.9	88.7	86.7	88.2	Config	pth [Link] (Access Code: 04OV)
RF-Learning total(Report)	94.1	88.6	94.9	94.5	82.4	82.0	82.6	-	-
RF-Learning total	94.5	90.0	94.0	94.1	81.5	82.0	84.7	Config	pth [Link] (Access Code: 49z1)

Visualization

Here is the picture for result visualization.

Citation

@article{rflearning,
  author={Hui Jiang and Yunlu Xu and Zhanzhan Cheng and Shiliang Pu and Yi Niu and Wenqi Ren and Fei Wu and Wenming Tan},
  title={Reciprocal Feature Learning via Explicit and Implicit Tasks in Scene Text Recognition},
  journal={CoRR},
  volume={abs/2105.06229},
  year={2021},
}

License

This project is released under the Apache 2.0 license

Contact

If there is any suggestion and problem, please feel free to contact the author with qiaoliang6@hikvision.com.