This is a re-implementation of the CRNN network, build by TensorFlow 2. This repository may help you to understand how to build an end-to-end text recognition network easily. By the way, here is the official repo implemented by bgshih.
This repo aims to build a simple, efficient, end-to-end text recognize network by using the various components of TensorFlow 2. The model build by the Keras functional API, the data pipeline build by tf.data
, and training with model.fit
, so we can use most of the functions provided by Tensorflow 2, such as Tensorboard
, Distribution strategy
, TensorFlow Profiler
etc.
tensorflow >= 2.2
Here I provide an example model that trained on the Mjsynth dataset, this model can only predict 0-9 and a-z(ignore case).
python demo.py -i example/images/ -t example/table.txt -m MODEL
Then, You will see output like this:
Path: example/images/1_Paintbrushes_55044.jpg, greedy: paintbrushes, beam search: paintbrushes
Path: example/images/3_Creationisms_17934.jpg, greedy: creationisms, beam search: creationisms
Path: example/images/2_Reimbursing_64165.jpg, greedy: reimbursing, beam search: reimbursing
Sometimes the beam search method will be better than the greedy method, but it's costly.
Before you start training, maybe you should prepare data first. The training process can visualize in Tensorboard. Because of that, we can check a lot of things, such as profile.
When image through the data input pipeline, the image shape will be resized to (32, width). The height is 32 because of CNN construction, the width will determine how many characters the model outputs at most.
All predictable characters are defined by the table.txt file, make sure the file is the same as the file used during training.
python train.py -ta TRAIN_ANNOTATION_PATHS -va VAL_ANNOTATION_PATHS -t TABLE_PATH
Before starting training, we can use restore
parameter to restore the model to convergence fastly, even if the number of characters is different.
For more parameters, please refer to the help.
Now, this repo support distributed training, just follow the official guide.
- Create an instance of the appropriate tf.distribute.Strategy.
- Move the creation of Keras model, optimizer and metrics inside strategy.scope.
Don't forget add drop_remainder=True
to tf.data.batch()
To train this network, you should prepare a lookup table, images and corresponding labels. Example data is copy from MJSynth dataset.
The file contains all characters and blank labels (in the last or any place both ok, but I find Tensorflow decoders can't change it now, so set it to last). By the way, you can write any word as blank.
It's an End-to-End method, so we don't need to indicate the position of the character in the image.
The labels corresponding to these three pictures are Paintbrushes
, Creationisms
, Reimbursing
.
We should write the image path and its corresponding label to a text file in a certain format such as example data. The data input pipeline will automatically detect the known format. Customization is also very simple, please check out the dataset.py
- MJSynth
- ICDAR2013
- [relative path] [label] such as [example.jpg awesome]
usage: eval.py [-h] -a ANNOTATION_PATHS [ANNOTATION_PATHS ...] -t TABLE_PATH
[-w IMAGE_WIDTH] [-b BATCH_SIZE] -m MODEL [--channels CHANNELS]
There are many components here to help us do other things. For example, deploy by Tensorflow serving
. Before you deploy, you can pick up a good weight, and convertes model to SavedModel
/h5
format by this command, it will add the Softmax layer in the last and cull the optimizer:
usage: converter.py [-h] -m MODEL -o OUTPUT
And now Tensorflow lite
also can convert this model, that means you can deploy it to Android, iOS etc.