Invertible Image Signal Processing

This repository includes official codes for "Invertible Image Signal Processing (CVPR2021)".

Figure: Our framework

Unprocessed RAW data is a highly valuable image format for image editing and computer vision. However, since the file size of RAW data is huge, most users can only get access to processed and compressed sRGB images. To bridge this gap, we design an Invertible Image Signal Processing (InvISP) pipeline, which not only enables rendering visually appealing sRGB images but also allows recovering nearly perfect RAW data. Due to our framework's inherent reversibility, we can reconstruct realistic RAW data instead of synthesizing RAW data from sRGB images, without any memory overhead. We also integrate a differentiable JPEG compression simulator that empowers our framework to reconstruct RAW data from JPEG images. Extensive quantitative and qualitative experiments on two DSLR demonstrate that our method obtains much higher quality in both rendered sRGB images and reconstructed RAW data than alternative methods.

Invertible Image Signal Processing
Yazhou Xing*, Zian Qian*, Qifeng Chen (* indicates joint first authors)
HKUST

[Paper] [Project Page] [Technical Video (Coming soon)]

Figure: Our results

Installation

Clone this repo.

git clone https://github.com/yzxing87/Invertible-ISP.git 
cd Invertible-ISP/

We have tested our code on Ubuntu 18.04 LTS with PyTorch 1.4.0, CUDA 10.1 and cudnn7.6.5. Please install dependencies by

conda env create -f environment.yml

Preparing datasets

We use MIT-Adobe FiveK Dataset for training and evaluation. To reproduce our results, you need to first download the NIKON D700 and Canon EOS 5D subsets from their website. The images (DNG) can be downloaded by

cd data/
bash data_preprocess.sh

The downloading may take a while. After downloading, we need to prepare the bilinearly demosaiced RAW and white balance parameters as network input, and ground truth sRGB (in JPEG format) as supervision.

python data_preprocess.py --camera="NIKON_D700"
python data_preprocess.py --camera="Canon_EOS_5D"

The dataset will be organized into

Training networks

We specify the training arguments into train.sh. Simply run

cd ../
bash train.sh

The checkpoints will be saved into ./exps/{exp_name}/checkpoint/.

Test and evaluation

To reconstruct the RAW from JPEG RGB, we need to first save the rendered RGB into disk then do test to recover RAW. Original RAW images are too huge to be directly tested on one 2080 Ti GPU. We provide two ways to test the model.

1. Subsampling the RAW for visualization purpose:

python test_rgb.py --task=EXPERIMENT_NAME \
                --data_path="./data/" \
                --gamma \
                --camera=CAMERA_NAME \
                --out_path=OUTPUT_PATH \
                --ckpt=CKPT_PATH

After finish, run

python test_raw.py --task=EXPERIMENT_NAME \
                --data_path="./data/" \
                --gamma \
                --camera=CAMERA_NAME \
                --out_path=OUTPUT_PATH \
                --ckpt=CKPT_PATH

2. Spliting the RAW data into patches, for quantitatively evaluation purpose. Turn on the --split_to_patch argument. See test.sh. The PSNR and SSIM metrics can be obtained by

python cal_metrics.py --path=PATH_TO_SAVED_PATCHES

Citation

@inproceedings{xing21invertible,
  title     = {Invertible Image Signal Processing},
  author    = {Xing, Yazhou and Qian, Zian and Chen, Qifeng},
  booktitle = {CVPR},
  year      = {2021}
}

Acknowledgement

Part of the codes benefit from DiffJPEG and Invertible-Image-Rescaling.

Contact

Free feel to contact me if there is any question. (Yazhou Xing, [email protected])