This is a c++ project deploying a deep scene text reading pipeline. It reads text from natural scene images.
The project is written in c++ using tensorflow computational framework. It is tested using tensorflow 1.4. Newer version should be ok too, but not tested. Please install:
-
Tensorflow
-
nsync project: https://github.com/google/nsync.git This is needed for building tensorflow.
-
opencv3.3
-
protobuf
-
eigen
Please check this project on how to build project using tensorflow with cmake: https://github.com/cjweeks/tensorflow-cmake It greatly helped the progress of building this project. When building tensorflow library, please be careful since we need to use opencv. Looks like there is still problem when including tensorflow and opencv together. It will make opencv unable to read image. Check out this issue: tensorflow/tensorflow#14267 The answer by allenlavoie solved my problem, so I paste it here:
"In the meantime, as long as you're not using any custom ops you can build libtensorflow_cc.so with bazel build --config=monolithic, which will condense everything together into one shared object (no libtensorflow_framework dependence) and seal off non-TensorFlow symbols. That shared object will have protocol buffer symbols."
Currently two pretrained model is provided. One for scene text detection, and one for scene text recognition. More model will be provided. Note that the current model is not so robust. U can easily change to ur trained model. The models will be continuously updated.
cd build
cmake ..
make
It will create an excutable named DetectText in bin folder.
The excutable could be excuted in three modes: (1) Detect (2) Recognize (3) Detect and Recognize
Download the pretrained detector model and put it in model/
./DetectText --detector_graph='model/Detector_model.pb'
--image_filename='test_images/test_img1.jpg' --mode='detect' --output_filename='results/output_image.jpg'
Download the pretrained recognizer model and put it in model/ Download the dictionary file and put it in model
./DetectText --recognizer_graph='model/Recognizer_model.pb'
--image_filename='test_images/recognize_image1.jpg' --mode='recognize'
--im_height=32 --im_width=128
Download the pretrained detector and recognizer model and put it in model/ as described previously.
./DetectText --recognizer_graph=$recognizer_graph --detector_graph='model/Detector_model.pb'
--image_filename='model/Recognizer_model.pb' --mode='detect_and_read' --output_filename='results/output_image.jpg'
-
Faster RCNN Detector Model The detector is trained with modified tensorflow [object detector api]: (https://github.com/tensorflow/models/tree/master/research/object_detection) I modify it by changing the proposal scheme to regress to the 4 coordinates of the oriented bounding box rather than regular rectangular bounding box. Check out this repo for the training code. Pretrained model: FasterRCNN_detector_model.pb
-
R2CNN will be updated. See R2CNN for details. The code is also modified with tnesorflow [object detector api]: (https://github.com/tensorflow/models/tree/master/research/object_detection) The training code will be released soon.
-
CTC scene text recognizer. The recognizer model follows the famous scene text recognition CRNN model
-
Spatial Attention OCR will be updated soon. It is based on GoogleOCR
The whole scene text reading pipeline detects the text and rotate it horizontally and read it with recognizer. The pipeline is here:
You can play with the code with provided pretrained models.
They are not fully optimized yet, but could be used for being familiar with the code.
Check them out here: models
You will find two detection models called: (1) FasterRCNN_detector_model.pb (2) R2CNN_detector_model.pb
Two recognition models with their charset: (1) Recognizer_model.pb + charset_full.txt and (2)Recognizer_model_case_insen.pb + charset_case_insen.txt.
Full charset means English letters + digit and case insen means case insensitive English letters + digit.
Let me know if u have any problens using them.
- Faster RCNN Faster RCNN paper.
- Tensorflow Object Detection API.
- An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition, reference paper for CRNN model.
- tensorflow-cmake, Tutorial of Building Project with tensorflow using cmake.
- R2CNN Reference paper for R2CNN.
- Dafang He. The Penn State University. [email protected] http://personal.psu.edu/duh188/