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++.

1. OpenCV 3.1 or above
2. CMake 2.8 or above
3. Visual Studio 2017 Community or above (windows only)

1. Install opencv, put opencv/ to C:\
2. add C:\opencv\build\x64\vc14\bin to System Variable Path
3. Use cmake to create Visual Studio Project

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

4. Compile via scene_text_recognition.sln, you will find executable in build/Release/

1. Install opencv, refer to OpenCV Installation in Linux
2. Use cmake to create Makefile, and make

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

Put scene_text_recognition, er_classifier/, ocr_classifier/, dictionary/ in the same directory.

./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  

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.

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%.

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.