真/伪4K图像数据集

针对学术界和产业界现有的4K超高清图像数据集只有真4K图像（即原始分辨率为4K的图像）而没有对应的由低分辨率图像上采样得到的伪4K图像这一问题，本课题提出了一个同时包含真4K图像和伪4K图像的数据集：SJTU-4K-2000。真/伪4K数据集包含真4K图像和伪4K图像。
为建立真4K内容数据集，我们首先从现有的4K数据集中（SJTU 4K，HDClub，NER center UHD）选取了21个4K视频序列。然后，我们从这些具有不同图像内容的视频序列中提取总共290张真4K图像，组成真4K内容数据集。真4K内容数据集包含的图像内容非常广泛，包括户外场景、室内场景、建筑、人物、动物、静物、夜景、运动场景、影视剧片段等。
为建立伪4K内容数据集，我们将所建立的真4K内容数据集的分辨率降至1080p(1920×1080像素)和720p(1280×720像素)。此外，我们在互联网上收集了大量1080p的影视剧视频序列。然后，我们通过不同的插值方法将它们上采样到4K分辨率。考虑到超分辨率(SR)算法的运行速度和增强效果之间的平衡，我们从传统超分辨算法、基于深度学习的超分辨率算法以及视频编辑软硬件的算法中选取了的14种常用的插值方法，分别是Bilinear、Bicubic、Lanczos2、Lanczos3、NEDI、DFDF、SRCNN、DSRCNN、ResNet、DResNet、Adobe Premiere、DaVinci Resolve 16、EditMax11、Imagine self-1uhd3上下转换器。所建立的伪4K内容数据集共有1825张伪4K图像，其中从1080p上采样得到的图像有1465张，从720p上采样得到的图像有360张。

Aiming at the limitation that existing 4K ultra-high-definition image datasets in both academia and industry only contain true 4K images (i.e., images with native 4K resolution) but lack corresponding pseudo 4K images generated by upsampling low-resolution counterparts, this study proposes a dataset named SJTU-4K-2000, which includes both true 4K and pseudo 4K images.

To construct the true 4K content dataset, we first selected 21 4K video sequences from existing public 4K datasets (SJTU 4K, HDClub, NER Center UHD). Subsequently, we extracted a total of 290 true 4K images from these video sequences covering diverse content to form the true 4K content dataset. The true 4K content dataset encompasses a wide spectrum of image content, including outdoor scenes, indoor scenes, architectures, human subjects, animals, still lifes, night scenes, motion scenarios, film and television clips, and more.

To develop the pseudo 4K content dataset, we downsampled the established true 4K content dataset to two low-resolution formats: 1080p (1920×1080 pixels) and 720p (1280×720 pixels). Furthermore, we collected a large corpus of 1080p film and television video sequences from the Internet. We then upsampled all these low-resolution images to 4K resolution using various interpolation methods. To balance the trade-off between inference speed and enhancement performance of super-resolution (SR) algorithms, we selected 14 commonly used interpolation methods from three categories: traditional super-resolution algorithms, deep learning-based super-resolution algorithms, and algorithms integrated in commercial video editing software and hardware. The selected methods are: Bilinear, Bicubic, Lanczos2, Lanczos3, NEDI, DFDF, SRCNN, DSRCNN, ResNet, DResNet, Adobe Premiere, DaVinci Resolve 16, EditMax11, and Imagine self-1uhd3 upscaler/downscaler.

The final pseudo 4K content dataset consists of 1825 pseudo 4K images in total, with 1465 images derived from upsampling 1080p content and 360 images generated by upsampling 720p content.

搜集汇总

背景与挑战

背景概述

该数据集是一个专门针对4K超高清图像的真伪对比数据集，旨在弥补现有数据集中缺乏伪4K图像的不足。它包含290张真4K图像和1825张伪4K图像，其中伪4K图像通过多种插值方法从低分辨率图像上采样生成，覆盖广泛场景，适用于图像处理和质量评估研究。数据集由上海交通大学创建，数据量达48.81GB，支持计算机视觉和超分辨率算法开发。