HsiehYiChia / Scene-text-recognition

Scene text detection and recognition based on Extremal Region(ER)

Scene text recognition

A real-time scene text recognition algorithm. Our system is able to recognize text in unconstrain background.
This algorithm is based on several papers, and was implemented in C/C++.

Enviroment and dependency

1. OpenCV 3.1 or above
2. CMake 3.10 or above
3. Visual Studio 2017 Community or above (Windows-only)

How to build?

Windows

1. Install OpenCV; put the opencv directory into C:\tools
   • You can install it manually from its Github repo, or
   • You can install it via Chocolatey: choco install opencv, or
   • If you already have OpenCV, edit CMakeLists.txt and change WIN_OPENCV_CONFIG_PATH to where you have it
2. Use CMake to generate the project files

   cd Scene-text-recognition
   mkdir build-win
   cd build-win
   cmake .. -G "Visual Studio 15 2017 Win64"

3. Use CMake to build the project

   cmake --build . --config Release

4. Find the binaries in the root directory

   cd ..
   dir | findstr scene

5. To execute the scene_text_recognition.exe binary, use its wrapper script; for example:

   .\scene_text_recognition.bat -i res\ICDAR2015_test\img_6.jpg

Linux

1. Install OpenCV; refer to OpenCV Installation in Linux
2. Use CMake to generate the project files

   cd Scene-text-recognition
   mkdir build-linux
   cd build-linux
   cmake ..

3. Use CMake to build the project

   cmake --build .

4. Find the binaries in the root directory

   cd ..
   ls | grep scene

5. To execute the binaries, run them as-is; for example:

   ./scene_text_recognition -i res/ICDAR2015_test/img_6.jpg

Usage

The executable file scene_text_recognition must ultimately exist in the project root directory (i.e., next to classifier/, dictionary/ etc.)

./scene_text_recognition -v:            take default webcam as input  
./scene_text_recognition -v [video]:    take a video as input  
./scene_text_recognition -i [image]:    take an image as input  
./scene_text_recognition -i [path]:     take folder with images as input,  
./scene_text_recognition -l [image]:    demonstrate "Linear Time MSER" Algorithm  
./scene_text_recognition -t detection:  train text detection classifier  
./scene_text_recognition -t ocr:        train text recognition(OCR) classifier 

Train your own classifier

Text detection

1. Put your text data to res/pos, non-text data to res/neg
2. Name your data in numerical, e.g. 1.jpg, 2.jpg, 3.jpg, and so on.
3. Make sure training folder exist
4. Run ./scene_text_recognition -t detection

mkdir training
./scene_text_recognition -t detection

5. Text detection classifier will be found at training folder

Text recognition(OCR)

1. Put your training data to res/ocr_training_data/
2. Arrange the data in [Font Name]/[Font Type]/[Category]/[Character.jpg], for instance Time_New_Roman/Bold/lower/a.jpg. You can refer to res/ocr_training_data.zip
3. Make sure training folder exist, and put svm-train to root folder (svm-train will be build by the system and should be found at build/)
4. Run ./scene_text_recognition -t ocr

mkdir training
mv svm-train scene-text-recognition/
scene_text_recognition -t ocr

5. Text recognition(OCR) classifier will be fould at training folder

How it works

The algorithm is based on an region detector called Extremal Region (ER), which is basically the superset of famous region detector MSER. We use ER to find text candidates. The ER is extracted by Linear-time MSER algorithm. The pitfall of ER is repeating detection, therefore we remove most of repeating ERs with non-maximum suppression. We estimate the overlapped between ER based on the Component tree. and calculate the stability of every ER. Among the same group of overlapped ER, only the one with maximum stability is kept. After that we apply a 2-stages Real-AdaBoost to fliter non-text region. We choose Mean-LBP as feature because it's faster compare to other features. The suviving ERs are then group together to make the result from character-level to word level, which is more instinct for human. Our next step is to apply an OCR to these detected text. The chain-code of the ER is used as feature and the classifier is trained by SVM. We also introduce several post-process such as optimal-path selection and spelling check to make the recognition result better.

Notes

For text classification, the training data contains 12,000 positive samples, mostly extract from ICDAR 2003 and ICDAR 2015 dataset. the negative sample are extracted from random images with a bootstrap process. As for OCR classification, the training data is consist of purely synthetic letters, including 28 different fonts.

The system is able to detect text in real-time(30FPS) and recognize text in nearly real-time(8~15 FPS, depends on number of texts) for a 640x480 resolution image on a Intel Core i7 desktop computer. The algorithm's end-to-end text detection accuracy on ICDAR dataset 2015 is roughly 70% with fine tune, and end-to-end recognition accuracy is about 30%.

Result

Detection result on IDCAR 2015

Recognition result on random image

Linear Time MSER Demo

The green pixels are so called boundry pixels, which are pushed into stacks. Each stack stand for a gray level, and pixels will be pushed according to their gary level.

References

1. D. Nister and H. Stewenius, “Linear time maximally stable extremal regions,” European Conference on Computer Vision, pages 183196, 2008.
2. L. Neumann and J. Matas, “A method for text localization and recognition in real-world images,” Asian Conference on Computer Vision, pages 770783, 2010.
3. L. Neumann and J. Matas, “Real-time scene text localization and recognition,” Computer Vision and Pattern Recognition, pages 35383545, 2012.
4. L. Neumann and J. Matas, “On combining multiple segmentations in scene text recognition,” International Conference on Document Analysis and Recognition, pages 523527, 2013.
5. H. Cho, M. Sung and B. Jun, ”Canny Text Detector: Fast and robust scene text localization algorithm,” Computer Vision and Pattern Recognition, pages 35663573, 2016.
6. B. Epshtein, E. Ofek, and Y. Wexler, “Detecting text in natural scenes with stroke width transform,” Computer Vision and Pattern Recognition, pages 29632970, 2010.
7. P. Viola and M. J. Jones, “Rapid object detection using a boosted cascade of simple features,” Computer Vision and Pattern Recognition, pages 511518, 2001.