pytorch>> Zero-DCE>> 返回
项目作者： bsun0802

项目描述 ：
Pytorch implementation of paper "Zero Reference low-light image enhancement"
高级语言： Python
项目地址: git://github.com/bsun0802/Zero-DCE.git
创建时间： 2020-05-02T16:19:51Z
项目社区：https://github.com/bsun0802/Zero-DCE
开源协议：
下载


Pytorch implementation of ZeroDCE

Link to the paper: Zero-Reference Deep Curve Estimation for Low-Light Image Enhancement

Some of my thoughts and observations during my implementatino journey of this non-reference image enhancement network can be found in Implementation Details.

This algorithm has potential for video enhancement. Here is a demo for applying vanilla ZeroDCE on a video clip without any improvements.

https://media.giphy.com/media/KkagL7vPeJ422010Z1/giphy.gif

If the video can not be displayed, see it here.

Inference with pre-trained model

I provide a pytorch model checkpoint that I trained, to use it:

  1. # go to the code directory
  2. cd code/
  4. # --testDir specify where the test images are stored
  5. python demo.py --device=-1 --testDir=../example-test-images/ \
  6.                --ckpt=../train-jobs/ckpt/train-jobs/ckpt/8LE-color-loss2_best_model.pth \
  7.                --output-dir=../demo-output

This will process the images in ../example-test-images/ and save results to ../demo-output. Demo results including output from ZeroDCE, and traditional gamma corrections for comparison.

Results

res_150_3

res_157_3

res_159_2

res_13_14

res_285_4

res_32_4

res_333_3

res_124_3

Comparison with traditional gamma correction

Please refer to comparison.pdf in demo-output/ for comparison of ZeroDCE and traditional gamma corrections. It shows result of train-jobs/ckpt/8LE-color-loss2_best_model.pth which is the best model when trainied with 160 epochs.

Results saved per 40 epochs can be downloaded here.

Generally, I think ~100 epochs should be good.

Training

Please note in order to use this repo, Python>=3.6 is required since I used f-strings.

Dataset

ZeroDCE is trained by SICE dataset, which can be downloaded here.

You can prepare the dataset for training with my code by modifying the paths in code/dataset.py to point to where you want to put the data.

Here is how the directory structure of the data should be in order to run my code. part1-512 is the root directory where I stored data.

  1. $ tree part1-512 --filelimit=10
  3. part1-512
  4. ├── test-toy
  5. │   ├── 318_3.JPG
  6. │   ├── 332_1.JPG
  7. │   ├── 340_1.JPG
  8. │   ├── 345_1.JPG
  9. │   ├── 353_3.JPG
  10. │   └── 356_7.JPG
  11. ├── train [2421 entries exceeds filelimit, not opening dir]
  12. └── val [600 entries exceeds filelimit, not opening dir]

Training and evaluation

I use relative path throughout my code, so please follow the exact directories structure as shown File Structure section.

Hyper-parameters are passed through command line and a dictionary named hp in train.py, for example:

  1. # go to the code directory
  2. cd code/
  5. # --experiment specify the prefix to use when storing outputs.
  6. python train.py --device=1 --numEpoch=120 --experiment=<@OUTPUT-PREFIX> --loss=1 \
  7.    --baseDir=../data/part1-512 --testDir=../data/test-1200-900 \
  8.    --weights 8 1.75 1 7 &
  10. # Note in train.py, STDOUT is directed to ../train-jobs/log/<PREFIX>.log, so if program raises errors, you need to find it there. 
  12. # train.py will call the evaluaton (eval.py) and store results to ../train-jobs/evaluation/ per 30 epoch.

To get detailed help on arguments, simply run python train.py --help or refers to the source code. Don’t be afraid, they should all be well-written and pythonic (suggestions are appreciated 👾).

File Structure

You need to follow this directory structure as I use relative paths. Upon root directory, you need to create

  • a code/ directory and put python files in it

  • a data/ directory and put subdirectory and data in it, considering modify dataset.py to your needs

  • directories train-jobs/log, train-jobs/ckpt, train-jobs/evaluation as log/checkpoing/results will be saved to them

image-20200503001251677

Training and implementation details

Here are some of my thoughts during implementation of this non-reference image enhancement work,

  1. The paper claim the parameter E in exposure loss, being in range [0.4, 0.7] does not cast a significant difference on the model. However, per my test E=0.6 works for this SICE dataset, all other E alwalys result in sub-optimal resutls. For example when E < 0.6, instead of increasing the darken pixel values, the model will decrease the pixels values that are more saturated (like white colored area and too bright area), resulting severe artifect in whitish objects and the border between dark and bright obejcts.
  2. My multiplier for Spatial Constancy Loss, Exposure Loss, Color Constancy Loss, and Total Variation Loss are 8, 1.75, 1, 8 respectively, this differ from the paper because the loss function implementation can be the same but off by a constant. For example, taking sum or taking mean, the loss are systematically the same but the magnitude are different,

Optimization of Original ZeroDCE

I devised a new version of ZeroDCE, which can achieve the same if not better results with 4-times of enhancement. The source code for this is under business license, if you want to know more about details. Shoot me an email bos AT usc DOT edu with subject ZeroDCE.

License

This project is licensed under the MIT License.

MIT © bsun