iq-scm / nhdrrnet

Keras Implementation of the paper Deep HDR Imaging via A Non-Local Network - TIP 2020

Deep HDR Imaging

The Keras Implementation of the Deep HDR Imaging via A Non-Local Network - TIP 2020

Content

• Deep-HDR-Imaging
• Getting Started
• Running
• References
• Citations

Getting Started

• Clone the repository

Prerequisites

• Tensorflow 2.2.0+
• Tensorflow_addons
• Python 3.6+
• Keras 2.3.0
• PIL
• numpy

pip install -r requirements.txt

Running

Training

• Preprocess

  • Download the training data and testing data.

  • Run this file to generate data. (Please remember to change path first)

  python src/create_dataset.py
  

• Train NHDRRNet

  python main.py
  

• Test NHDRRNet

  python test.py
  

Usage

Training

usage: main.py [-h] [--images_path IMAGES_PATH] [--test_path TEST_PATH]
               [--lr LR] [--gpu GPU] [--num_epochs NUM_EPOCHS] 
               [--train_batch_size TRAIN_BATCH_SIZE]
               [--display_ep DISPLAY_EP] [--checkpoint_ep CHECKPOINT_EP]
               [--checkpoints_folder CHECKPOINTS_FOLDER]
               [--load_pretrain LOAD_PRETRAIN] [--pretrain_dir PRETRAIN_DIR]
               [--filter FILTER] [--kernel KERNEL]
               [--encoder_kernel ENCODER_KERNEL]
               [--decoder_kernel DECODER_KERNEL]
               [--triple_pass_filter TRIPLE_PASS_FILTER]

optional arguments: -h, --help                show this help message and exit
                    --images_path             training path
                    --lr                      LR
                    --gpu                     GPU
                    --num_epochs              NUM of EPOCHS
                    --train_batch_size        training batch size
                    --display_ep              display result every "x" epoch
                    --checkpoint_ep           save weights every "x" epoch
                    --checkpoints_folder      folder to save weight
                    --load_pretrain           load pretrained model
                    --pretrain_dir            pretrained model folder
                    --filter                  default filter
                    --kernel                  default kernel
                    --encoder_kernel          encoder filter size
                    --decoder_kernel          decoder filter size
                    --triple_pass_filter      number of filter in triple pass

Testing

The weight file was deprecated. Will be updated soon.

usage: test.py [-h] [--test_path TEST_PATH] [--gpu GPU]
                    [--weight_test_path WEIGHT_TEST_PATH] [--filter FILTER]
                    [--kernel KERNEL] [--encoder_kernel ENCODER_KERNEL]
                    [--decoder_kernel DECODER_KERNEL]
                    [--triple_pass_filter TRIPLE_PASS_FILTER]

optional arguments: -h, --help                    show this help message and exit
                    --test_path                   test path
                    --weight_test_path            weight test path
                    --filter                      default filter
                    --kernel                      default kernel
                    --encoder_kernel              encoder filter size
                    --decoder_kernel              decoder filter size
                    --triple_pass_filter          number of filter in triple pass

Result

License

This project is licensed under the MIT License - see the LICENSE file for details

References

[1] Deep HDR Imaging via A Non-Local Network - TIP 2020 link

[3] Training and Testing dataset - link

Citation

    @ARTICLE{8989959,  author={Q. Yan and L. Zhang and Y. Liu and Y. Zhu and J. Sun and Q. Shi and Y. Zhang},  
    journal={IEEE Transactions on Image Processing},   
    title={Deep HDR Imaging via A Non-Local Network},   
    year={2020},  
    volume={29},  
    number={},  
    pages={4308-4322},}

Acknowledgments

• This work based on the paper mentioned above with few modification:
  • the fixed size of the adaptive average pooling (16 instead of 32 as assigned in the paper)
  • the number of triple pass module is defined as 10 to match the number of 32M as stated in the paper.
• Any ideas on updating or misunderstanding, please send me an email: [email protected]
• If you find this repo helpful, kindly give me a star.

Update: I have just released my work on HDR imaging using Attention non-local network. Please check as follow: https://github.com/tuvovan/ANL-HDRI