PyTorch code for 2023 paper "Raw Image Based Over-Exposure Correction Using Channel-Guidance Strategy"
-
We present the first benchmark to explore the superiority of RAW images based on a detailed analysis in the over-exposure correction. Building upon these insights, we have developed a novel end-to-end RAW-to-sRGB network that leverages a data-driven approach to address the challenges of real-world over-exposure photography.
-
We develop a new RAW-based non-green channel guidance strategy to maximize the utilization of useful information from red and blue channels and exploit the effective way to use it.
-
We have assembled a new real-world dataset specifically designed for over-exposure correction. This dataset encompasses both RAW and sRGB images across a broad spectrum of scenes. Quantitative and qualitative results on both synthetic and real-world datasets show that our CGNet achieves state-of-the-art performance on overexposure correction.
We provide Prerequisites for reference, please refer to requirements.txt.
- matplotlib==3.5.3
- numpy==1.21.6
- opencv-python==4.7.0.72
- Pillow==9.4.0
- rawpy==0.18.0
- scipy==1.7.3
- torch==1.10.0+cu111
- torchvision==0.11.0+cu111
- tqdm==4.65.0
Over-exposed Raw image processing has been rarely studied due to limited available data. In order to bridge the gap of datasets and make it feasible for RAW-based end-to-end learning, we construct a large-scale RAW-based synthetic dataset mainly for model pretraining, and collect a real-world dataset from real photography that contains diverse over-exposed image pairs for training, fine-tuning and evaluation. Both of them are created in both RAW and sRGB formats and contain paired over-exposed and properly-exposed images.
To simulate realistic multi-exposure errors, we render over-exposed RAW images by multiplying reference RAW images with 4 different digital ratios of 3, 5, 8, and 10. 1595 high-quality reference original images are finally retained in our SOF dataset. Each properly-exposed reference image corresponds to 4 overexposed images of different degrees.Since the original data files are difficult to verify, the synthetic data we finally collected (processed from the MIT-Adobe FiveK dataset) totaled 3051 groups. During the test process, the program obtains the corresponding test data by reading the txt file.
For more information ,please click here
The complete Synthetic RAW Image Dataset (~94.31GB) is available via link ：https://pan.baidu.com/s/14sm4ePAr2xBf442hmjmqlA Extraction code：kfkg
The collected Real-world Paired Over-exposure (RPO) dataset contains 650 indoor and outdoor scenarios. For each scene, we collect a sequence of overexposed RAW images with 4 pre-set over-exposure ratios to evaluate over-exposure correction methods. This yields a total of 2,600 over-exposure RAW images, with the corresponding sRGB images.
For more information ,please click [here](https://github.com/whiteknight-WJN/SOF-Dataset
The complete Real-World RAW Image Dataset (~22.32GB) is available via link：https://pan.baidu.com/s/1L6Fog7X6Xd3tD_aRsyVNtg Extraction code：stzm
Checkpoint is available via link: https://pan.baidu.com/s/1WtT128RYSja12ogVXziX7Q Extraction code：nw3a
If you find our work helpful to your research or work, Please cite our paper.
@ARTICLE{10239166,
author={Fu, Ying and Hong, Yang and Zou, Yunhao and Liu, Qiankun and Zhang, Yiming and Liu, Ning and Yan, Chenggang},
journal={IEEE Transactions on Circuits and Systems for Video Technology},
title={Raw Image Based Over-Exposure Correction Using Channel-Guidance Strategy},
year={2023},
volume={},
number={},
pages={1-1},
doi={10.1109/TCSVT.2023.3311766}}