智能手机图像去噪训练数据集（SIDD）

在过去的十年里，从单反相机和傻瓜相机的成像到智能手机相机的成像发生了天文数字的转变。由于小光圈和传感器尺寸，智能手机图像的噪声明显高于单反图像。虽然智能手机图像的去噪是一个活跃的研究领域，但研究界目前缺乏一个去噪图像数据集，该数据集代表了来自具有高质量地面实况的智能手机相机的真实噪声图像。我们在本文件中对这一问题作出了以下贡献。我们提出了一种系统的方法来估计有噪声图像的真实情况，该方法可用于衡量智能手机相机的去噪性能。使用该程序，我们使用五台具有代表性的智能手机相机，从不同照明条件下的10个场景中捕获了约30000幅噪声图像的数据集，即智能手机图像去噪数据集（SIDD），并生成了它们的真实图像。我们使用这个数据集对许多去噪算法进行了基准测试。我们表明，在我们的高质量数据集上训练时，基于CNN的方法比使用替代策略（如用作地面实况数据代理的低ISO图像）训练时表现更好。 Papers Abdelrahman Abdelhamed，Lin S.，Brown M.S.“智能手机摄像头的高质量去噪数据集”，IEEE计算机视觉和模式识别（CVPR），2018年6月。 [PDF] [Bibtex] Abdelrahman Abdelhamed，Timofte R.，Brown M.S.等人，“NTIRE 2019真实图像去噪挑战：方法和结果”，IEEE计算机视觉和模式识别研讨会（CVPRW），2019年6月。[PDF] [Bibtex] Code Ground-truth image estimation A simple camera pipeline for rendering raw-RGB images into sRGB. License The dataset and the associated code repositories are under the MIT License. Contact For any questions, remarks, or comments, please contact: Abdelrahman Abdelhamed.

Over the past decade, imaging from DSLR and point-and-shoot cameras has undergone a dramatic shift toward smartphone camera imaging. Due to their small aperture and sensor size, smartphone images exhibit significantly higher noise levels than DSLR images. While smartphone image denoising is an active research area, the research community currently lacks a denoising image dataset that represents real noisy images captured by smartphone cameras with high-quality ground truth. We make the following contributions to address this issue in this document: We propose a systematic method to estimate ground truth for noisy images, which can be used to evaluate the denoising performance of smartphone cameras. Using this procedure, we captured a dataset of approximately 30,000 noisy images across 10 scenes under varying lighting conditions using five representative smartphone cameras, and generated their corresponding ground truth images. This dataset is named the Smartphone Image Denoising Dataset (SIDD). We use this dataset to benchmark numerous denoising algorithms. We demonstrate that CNN-based methods trained on our high-quality dataset perform better than those trained using alternative strategies, such as low-ISO images used as proxies for ground truth data.
Papers
Abdelrahman Abdelhamed, Lin S., Brown M. S. "High-Quality Denoising Dataset for Smartphone Cameras", IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. [PDF] [Bibtex]
Abdelrahman Abdelhamed, Timofte R., Brown M. S. et al., "NTIRE 2019 Real Image Denoising Challenge: Methods and Results", IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), June 2019. [PDF] [Bibtex]
Code
Ground-truth image estimation
A simple camera pipeline for rendering raw-RGB images into sRGB.
License
The dataset and the associated code repositories are licensed under the MIT License.
Contact
For any questions, remarks, or comments, please contact: Abdelrahman Abdelhamed.

搜集汇总

背景与挑战

背景概述

智能手机图像去噪训练数据集（SIDD）是一个专门用于智能手机图像去噪研究的高质量数据集，包含约30000幅来自五台代表性智能手机相机的噪声图像，覆盖10个不同照明场景，并提供了真实图像作为地面实况。该数据集用于评估去噪算法性能，研究表明基于CNN的方法在其上训练时表现优异，相关论文发表于IEEE CVPR等会议，许可证为MIT License。