This is a TensorFlow implementation of a Deep Neural Network for scene text recognition. It is 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 the paper for architecture details. Thanks to the author Baoguang Shi.

The model consists of a CNN stage extracting features which are fed to an RNN stage (Bi-LSTM) and a CTC loss.

This software has been developed on Ubuntu 16.04(x64) using python 3.5 and TensorFlow 1.12. Since it uses some recent features of TensorFlow it is incompatible with older versions.

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 you will find a model pre-trained on the Synth 90kdataset. When the tfrecords file of synth90k dataset has been successfully generated you may evaluated the model by the following script

The pretrained crnn model weights on Synth90k dataset can be found here

python tools/evaluate_shadownet.py --dataset_dir PATH/TO/YOUR/DATASET_DIR 
--weights_path PATH/TO/YOUR/MODEL_WEIGHTS_PATH
--char_dict_path PATH/TO/CHAR_DICT_PATH 
--ord_map_dict_path PATH/TO/ORD_MAP_PATH
--process_all 1 --visualize 1

If you set visualize true the expected output during evaluation process is

After all the evaluation process is done you should see some thing like this:

The model's main evaluation index are as follows:

Test Dataset Size: 891927 synth90k test images

Per char Precision: 0.974325 without average weighted on each class

Full sequence Precision: 0.932981 without average weighted on each class

For Per char Precision:

single_label_accuracy = correct_predicted_char_nums_of_single_sample / single_label_char_nums

avg_label_accuracy = sum(single_label_accuracy) / label_nums

For Full sequence Precision:

single_label_accuracy = 1 if the prediction result is exactly the same as label else 0

avg_label_accuracy = sum(single_label_accuracy) / label_nums

Part of the confusion matrix of every single char looks like this:

If you want to test a single image you can do it with

python tools/test_shadownet.py --image_path PATH/TO/IMAGE 
--weights_path PATH/TO/MODEL_WEIGHTS
--char_dict_path PATH/TO/CHAR_DICT_PATH 
--ord_map_dict_path PATH/TO/ORD_MAP_PATH

Example test_01.jpg

Example test_02.jpg

Example test_03.jpg

Download the whole synth90k dataset here And extract all th files into a root dir which should contain several txt file and several folders filled up with pictures. Then you need to convert the whole dataset into tensorflow records as follows

python tools/write_text_features 
--dataset_dir PATH/TO/SYNTH90K_DATASET_ROOT_DIR
--save_dir PATH/TO/TFRECORDS_DIR

During converting all the source image will be scaled into (32, 100)

For all the available training parameters, check global_configuration/config.py, then train your model with

python tools/train_shadownet.py --dataset_dir PATH/TO/YOUR/TFRECORDS
--char_dict_path PATH/TO/CHAR_DICT_PATH 
--ord_map_dict_path PATH/TO/ORD_MAP_PATH

If you wish, you can add more metrics to the training progress messages with --decode_outputs 1, but this will slow training down. You can also continue the training process from a snapshot with

python tools/train_shadownet.py --dataset_dir PATH/TO/YOUR/TFRECORDS
--weights_path PATH/TO/YOUR/PRETRAINED_MODEL_WEIGHTS
--char_dict_path PATH/TO/CHAR_DICT_PATH --ord_map_dict_path PATH/TO/ORD_MAP_PATH

If you has multiple gpus in your local machine you may use multiple gpu training to access a larger batch size input data. This will be supported as follows

python tools/train_shadownet.py --dataset_dir PATH/TO/YOUR/TFRECORDS
--char_dict_path PATH/TO/CHAR_DICT_PATH --ord_map_dict_path PATH/TO/ORD_MAP_PATH
--multi_gpus 1

The sequence distance is computed by calculating the distance between two sparse tensors so the lower the accuracy value is the better the model performs. The training accuracy is computed by calculating the character-wise precision between the prediction and the ground truth so the higher the better the model performs.

You can convert the ckpt model into tensorflow saved model for tensorflow service by running following script

bash tools/export_crnn_saved_model.sh

The original experiment run for 2000000 epochs, with a batch size of 32, an initial learning rate of 0.01 and exponential decay of 0.1 every 500000 epochs. During training the train loss dropped as follows

The val loss dropped as follows

I have uploaded a newly trained crnn model on chinese dataset which can be found here. Sorry for not knowing the owner of the dataset. But thanks for his great work. If someone knows it you're welcome to let me know. The pretrained weights can be found here

Before start training you may need reorgnize the dataset's label information according to the synth90k dataset's format if you want to use the same data feed pip line mentioned above. Now I have reimplemnted a more efficient tfrecords writer which will accelerate the process of generating tfrecords file. You may refer to the code for details. Some information about training is listed bellow:

image size: (280, 32)

classes nums: 5824 without blank

sequence length: 70

training sample counts: 2733004

validation sample counts: 364401

testing sample counts: 546601

batch size: 32

training iter nums: 200000

init lr: 0.01

Example test_01.jpg

Example test_02.jpg

Example test_03.jpg

The val loss dropped as follows

Add a small demo to recognize chinese pdf using the chinese crnn model weights. If you want to have a try you may follow the command:

cd CRNN_ROOT_REPO
python tools/recongnize_chinese_pdf.py -c ./data/char_dict/char_dict_cn.json 
-o ./data/char_dict/ord_map_cn.json --weights_path model/crnn_chinese/shadownet.ckpt 
--image_path data/test_images/test_pdf.png --save_path pdf_recognize_result.txt

You should see the same result as follows:

The left image is the recognize result displayed on console and the right image is the origin pdf image.

The left image is the recognize result written in local file and the right image is the origin pdf image.

• Add new model weights trained on the whole synth90k dataset
• Add multiple gpu training scripts
• Add new pretrained model on chinese dataset
• Add an online toy demo
• Add tensorflow service script