This is a tensorflow re-implementation of PSENet: Shape Robust Text Detection with Progressive Scale Expansion Network.

Thanks for the author's (@whai362) awesome work!

1. Any version of tensorflow version > 1.0 should be ok.
2. python 2 or 3 will be ok.

trained on ICDAR 2015 (training set) + ICDAR2017 MLT (training set):

baiduyun extract code: pffd

google drive

This model is not as good as article's, it's just a reference. You can finetune on it or you can do a lot of optimization based on this code.

If you want to train the model, you should provide the dataset path, in the dataset path, a separate gt text file should be provided for each image, and make sure that gt text and image file have the same names.

Then run train.py like:

python train.py --gpu_list=0 --input_size=512 --batch_size_per_gpu=8 --checkpoint_path=./resnet_v1_50/ \
--training_data_path=./data/ocr/icdar2015/

If you have more than one gpu, you can pass gpu ids to gpu_list(like --gpu_list=0,1,2,3)

Note:

1. right now , only support icdar2017 data format input, like (116,1179,206,1179,206,1207,116,1207,"###"), but you can modify data_provider.py to support polygon format input
2. Already support polygon shrink by using pyclipper module
3. this re-implementation is just for fun, but I'll continue to improve this code.
4. re-implementation pse algorithm by using c++ (if you use python2, just run it, if python3, please replace python-config with python3-config in makefile)

run eval.py like:

python eval.py --test_data_path=./tmp/images/ --gpu_list=0 --checkpoint_path=./resnet_v1_50/ \
--output_dir=./tmp/

a text file and result image will be then written to the output path.

If you encounter any issue check issues first, or you can open a new issue.

1. https://download.tensorflow.org/models/resnet_v1_50_2016_08_28.tar.gz
2. https://github.com/CharlesShang/FastMaskRCNN
3. whai362/PSENet#15
4. https://github.com/argman/EAST

@rkshuai found a bug about concat features in model.py.

If this repository helps you，please star it. Thanks.