Skip to content

Latest commit

 

History

History
132 lines (107 loc) · 5.05 KB

README.md

File metadata and controls

132 lines (107 loc) · 5.05 KB
Raw

Zero-DCE TF

The Tensorflow Implementation of the Zero-Reference Deep Curve Estimation for Low-Light Image Enhancement - CVPR 2020

Update:

I pushed my project to Google Cloud Platform. May need more improvement. Should you have any comment or inquiries or just basically want to enhance your images, give it a try here

Content

Getting Started

  • Clone the repository

Prerequisites

  • Tensorflow 2.2.0+
  • Python 3.6+
  • Keras 2.3.0
  • PIL
  • numpy
pip install -r requirements.txt

Running

Training

  • Preprocess

    • Download the training data at Google Drive.

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

    python src/prepare_data.py

  • Train ZERO_DCE

    python train.py

  • Test ZERO_DCE

    python test.py

Usage

Training

python train.py [-h] [--lowlight_images_path LOWLIGHT_IMAGES_PATH] [--lr LR]
                [--num_epochs NUM_EPOCHS] [--train_batch_size TRAIN_BATCH_SIZE]
                [--val_batch_size VAL_BATCH_SIZE] [--display_iter DISPLAY_ITER]
                [--checkpoint_iter CHECKPOINT_ITER] [--checkpoints_folder CHECKPOINTS_FOLDER]
                [--load_pretrain LOAD_PRETRAIN] [--pretrain_dir PRETRAIN_DIR]

optional arguments: -h, --help                show this help message and exit
                    --lowlight_images_path    LOWLIGHT_IMAGES_PATH
                    --lr                      LR
                    --num_epochs              NUM_EPOCHS
                    --train_batch_size        TRAIN_BATCH_SIZE
                    --val_batch_size          VAL_BATCH_SIZE
                    --display_iter            DISPLAY_ITER
                    --checkpoint_iter         CHECKPOINT_ITER
                    --checkpoints_folder      CHECKPOINTS_FOLDER
                    --load_pretrain           LOAD_PRETRAIN
                    --pretrain_dir            PRETRAIN_DIR

Testing

python test.py [-h] [--lowlight_test_image_path]

optional arguments: -h, --help                    show this help message and exit
                    --lowlight_test_image_path    LOWLIGHT_TEST_IMAGES_PATH

Video-DCE

Video-DCE is a simple adaptation of the lowlight image enhancement script to take videos as input, processing individual frames with the model, then outputing a HuffYUV encoded video while copying the existing audio track from the input video. Unlike the image processing scripts, Video-DCE will not resize the input video frames in order to process them. The output video will mimic the input's video FPS, but you can specify a different Display Aspect Ratio if needed.

usage: video-dce.py [-h] --input_video INPUT_VIDEO [--output_video OUTPUT_VIDEO] [--max_frames MAX_FRAMES] [--dar DAR]

MAX_FRAMES can be useful for testing results out of a portion of the video. Frame counting will always start from zero.

The script has only been tested with SD resolution videos (720x486, 640x480 and 720x480) as it's my main use case, so there might be bugs depending on the resolution of your input video.

Result

INPUT OUTPUT
INPUT OUTPUT
INPUT OUTPUT
INPUT OUTPUT
INPUT OUTPUT
input output

License

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

References

[1] Zero-Reference Deep Curve Estimation for Low-Light Image Enhancement - CVPR 2020 link

[3] Low-light dataset - link

Citation

    @misc{guo2020zeroreference,
    title={Zero-Reference Deep Curve Estimation for Low-Light Image Enhancement},
    author={Chunle Guo and Chongyi Li and Jichang Guo and Chen Change Loy and Junhui Hou and Sam Kwong and Runmin Cong},
    year={2020},
    eprint={2001.06826},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

Try on web

The project is now available on GCP. Give it a try

Acknowledgments

  • This repo is the re-production of the original pytorch version
  • Thank you for helping me to understand more about pains that tensorflow may cause.
  • Final words:
    • Any ideas on updating or misunderstanding, please send me an email: [email protected]
    • If you find this repo helpful, kindly give me a star.