Shape Robust Text Detection with Progressive Scale Expansion Networkt

A reimplement of PSENet with tensorflow. Only trained on IC15,TD500,and CTW1500. The preformence is low compared to paper's result. Maybe because of partial data training (For each dataset I don't use extra data, but the paper use MLT data), or just maybe because of some bug in this code.

• The pse algrithm is implemented with C++ and warped by cython in floder in ./PSE_C
• Support simple hype-parameter search for test(such as threshold)
• Support checkpoint restore.
• summary logging by tensorboard

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

Prerequisites

pip install -r requriment.txt

• tensorflow = 1.8
• pudb(for debug)
• pyclipper(to clip polygon)
• skimage
• Polygon3
• Cython

Result(Train only with the corresponding dataset, no extra data)

- this model had been removed in my server, so I forget exact f-measure, just for reference

Training

1. Converting the dataset to tfrecord for high performence reading. This converter support IC15,TD500,CTW1500. Default for IC15. And you need to change dataset_dir path in dataset_factory.py.

    cd dataset
    python write_tfrecord.py --data-folder 'your path here'
   

2. To use ImageNet Pretrained model, download the model to Logs/model from model
3. Train the model, you need edit configuration.py to adjust the patameter of trianing such as bactch size,learning rate.

   mkdir Logs/train
   python train_PSENet.py --run_name test --restore=True --use_pretrain=True --gpus '0' --ss=10 --se=10 --about=''
   

4. you can monitor the progress of training by tensorboard

   tensorboard --host localhost --samples_per_plugin images=500 --port 7000 --logdir Logs/train/test
   

   

Testing

1. Compile the PSE code

   python setup.py build_ext --inplace
   

2. Edit test config in configuration.py such as threshold, image size, test_dir is the path of your test images.(such as /you/path/icdar2015/ch4_test_images)

   python eval_metric.py --ma=True --train_name test --gpus='2' --lg=False
   

3. If lg is True, you can view detect result in folder Logs/test/<run_name>/model.ckpt-399999_0/image_log