智能手机上从RGB重建RAW图像：NTIRE 2025挑战赛报告
图像与视频处理
2025-06-03 v1 计算机视觉与模式识别
摘要
由于RAW域具有线性特性、位深和传感器设计等优势，许多低级视觉任务在RAW域中进行。尽管如此，RAW图像数据集仍然稀缺，且收集成本高于已有的庞大全公开sRGB数据集。因此，许多方法尝试利用传感器信息和sRGB图像生成逼真的RAW图像。本文涵盖了第二届从sRGB重建RAW（逆向ISP）挑战赛。我们的目标是在没有元数据的情况下，根据对应的sRGB图像恢复智能手机的RAW传感器图像，从而“逆转”ISP变换。超过150名参与者加入了此次NTIRE 2025挑战赛并提交了高效的模型。所提出的方法和基准测试为生成逼真的RAW数据确立了最新技术水平。
关键词
引用
@article{arxiv.2506.01947,
title = {RAW Image Reconstruction from RGB on Smartphones. NTIRE 2025 Challenge Report},
author = {Marcos V. Conde and Radu Timofte and Radu Berdan and Beril Besbinar and Daisuke Iso and Pengzhou Ji and Xiong Dun and Zeying Fan and Chen Wu and Zhansheng Wang and Pengbo Zhang and Jiazi Huang and Qinglin Liu and Wei Yu and Shengping Zhang and Xiangyang Ji and Kyungsik Kim and Minkyung Kim and Hwalmin Lee and Hekun Ma and Huan Zheng and Yanyan Wei and Zhao Zhang and Jing Fang and Meilin Gao and Xiang Yu and Shangbin Xie and Mengyuan Sun and Huanjing Yue and Jingyu Yang Huize Cheng and Shaomeng Zhang and Zhaoyang Zhang and Haoxiang Liang},
journal= {arXiv preprint arXiv:2506.01947},
year = {2025}
}
备注
CVPR 2025 - New Trends in Image Restoration and Enhancement (NTIRE)