unparied night-to-day translation

Introduction

This code is for night-to-day translation. Image processing codes is in the directory "process". In addition to the paper, we made some attempts at image sharpening and handling local over-exposure. The codes is developed by HollyLee, RuiZhu et al. And the work is based on many previous studies, To be specific:
Cycle-GAN paper：https://arxiv.org/pdf/1703.10593.pdf, by YanJun Zhu et al.
pix2pix paper: https://arxiv.org/pdf/1611.07004.pdf, by Phillip Isola et al.
To improve training of wasserstein GANs: https://arxiv.org/abs/1704.00028, by Gulrajani et al.
To construct resnet generator, we adapt torch code and idea from Justin Johnson's neural style transfer project(https://github.com/jcjohnson/fast-neural-style) And their paper: https://arxiv.org/pdf/1512.03385.pdf.
DCP paper: http://kaiminghe.com/publications/cvpr09.pdf
We adapt the idea of self-attention map from Enlighten-GAN: https://arxiv.org/abs/1906.06972

Run the code for unpaired mode (only to generate pictures)

Put your images from night domain in ./process/in_img and run ./process/run.py for enhancing.

Images after enhancement will be stored in ./process/out_img and dehazing results of those enhanced images are in ./process/final_out_img.

To achieve stlye-transfer, put yor enhanced night images(not dehazed) in ./datasets/enhanced2daylight/trainA and daylight images in ./datasets/enhanced2daylight/trainB

• Train the model:

python train.py --dataroot ./datasets/enhanced2daylight --name enhanced2daylight_cyclegan --model cycle_gan

• To view training results and loss plots, run python -m visdom.server and click the URL http://localhost:8097.

  To see more intermediate results, check out ./checkpoints/enhanced2daylight_cyclegan/web/index.html.

  To test the model, you're supposed to put your test images in ./datasets/enhanced2daylight/testA for enhanced images and ./datasets/enhanced2daylight/testB for daylight images

• Test the model for translating both sides(always not necessary):

python test.py --dataroot ./datasets/enhanced2daylight --name enhanced2daylight_cyclegan --model cycle_gan

• Test the model for translating from only one side(e.g. from enhanced to daylight):

python test.py --dataroot datasets/enhanced2daylight/testA --name enhanced2daylight_cyclegan --model test --no_dropout

• The test results will be saved here: ./results/enhanced2daylight_cyclegan/latest_test. After style transfer, do the haze removal if necessary. Put hazed images in ./process/defog_in and run ./process/run_defog.py The final result will be in ./process/defog_out

##supplementary specification In directory./process, there are some codes for image processing:
batch_rename.py: Batch rename image files in the folder.
run_clearify.py: Batch clearify image files in the folder.
run.py: Batch process image files by Retinex in the folder. (and we adopt MSRCP)
run_defog.py: Batch defog image files in the folder.
HLS.py: Batch adjust brightness and contrast of image files in the folder.
HLS_overexposure.py: some attempts to solve over-exposure.
proVideo.py: separates video frames and sounds.
make_video.py: make video.
find_dif.py: Restoration of active objects erased by overfitting.