Use tensorflow to implement a Deep Neural Network for scene text recognition mainly based on the paper "An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition".You can refer to their paper for details http://arxiv.org/abs/1507.05717. Thanks for the author Baoguang Shi.
This model consists of a CNN stage, RNN stage and CTC loss for scene text recognition task.
This software has mostly been tested on Ubuntu 16.04(x64) using python3.5 and tensorflow 1.3.0 with cuda-8.0, cudnn-6.0 and a GTX-1070 GPU. Other versions of tensorflow have not been tested but might work properly above version 1.0.
The following methods are provided to install dependencies:
There are Dockerfiles inside the folder docker. Follow the instructions inside docker/README.md to build the images.
You can create a conda environment with the required dependencies using:
conda env create -f crnntf-env.yml
Required packages may be installed with
pip3 install -r requirements.txt
In this repo I uploaded a model trained on a subset of the Synth 90k. During data preparation process the dataset is converted into a tensorflow records which you can find in the data folder. You can test the trained model on the converted dataset by
python tools/test_shadownet.py --dataset_dir data/ --weights_path model/shadownet/shadownet_2017-09-29-19-16-33.ckpt-39999
Expected output is
If you want to test a single image you can do it by
python tools/demo_shadownet.py --image_path data/test_images/test_01.jpg --weights_path model/shadownet/shadownet_2017-09-29-19-16-33.ckpt-39999
Example image_01 is
Expected output_01 is
Example image_02 is
Expected output_02 is
Example image_03 is
Expected output_03 is
Example image_04 is
Expected output_04 is
Firstly you need to store all your image data in a root folder then you need to supply a txt file named sample.txt to specify the relative path to the image data dir and it's corresponding text label. For example
path/1/2/373_coley_14845.jpg coley
path/17/5/176_Nevadans_51437.jpg nevadans
Secondly you are supposed to convert your dataset into tensorflow records which can be done by
python tools/write_text_features --dataset_dir path/to/your/dataset --save_dir path/to/tfrecords_dir
All your training image will be scaled into (32, 100, 3) the dataset will be divided into train, test, validation set and you can change the parameter to control the ratio of them.
The whole training epoches are 40000 in my experiment. I trained the model with a batch size 32, initialized learning rate is 0.1 and decrease by multiply 0.1 every 10000 epochs. For more training parameters information you can check the global_configuration/config.py for details. To train your own model by
python tools/train_shadownet.py --dataset_dir path/to/your/tfrecords
You can also continue the training process from the snapshot by
python tools/train_shadownet.py --dataset_dir path/to/your/tfrecords --weights_path path/to/your/last/checkpoint
After several times of iteration you can check the log file in logs folder you are supposed to see the following contenent
The seq distance is computed by calculating the distance between two saparse tensor so the lower the accuracy value is the better the model performs.The train accuracy is computed by calculating the character-wise precision between the prediction and the ground truth so the higher the better the model performs.
During my experiment the loss drops as follows
The distance between the ground truth and the prediction drops as follows
The accuracy during training process rises as follows
The model is trained on a subet of Synth 90k. So i will train a new model on the whold dataset to get a more robust model.The crnn model needs large of training data to get a rubust model.