RealSRSet

20张真实的低分辨率图像，选自现有数据集或从互联网下载；通过添加不同噪声级别的高斯噪声，采用不同质量因子的JPEG压缩，并通过反向正向摄像机图像信号处理生成处理后的摄像机传感器噪声来合成噪声(ISP) 管道模型和 RAW 图像噪声模型。为了验证新退化模型的有效性，我们训练了一个深度盲 ESRGAN 超级解析器，然后将其应用于具有不同退化的合成图像和真实图像的超级解析。实验结果表明，新的退化模型有助于显着提高深度超解算器的实用性，从而为实际 SISR 应用提供强大的替代解决方案。

Twenty real low-resolution images were selected from existing datasets or downloaded from the public Internet. To synthesize noise for the camera Image Signal Processing (ISP) pipeline model and RAW image noise model, we applied three degradation strategies: adding Gaussian noise with different noise levels, performing JPEG compression with varying quality factors, and generating processed camera sensor noise via the inverse forward camera image signal processing pipeline. To validate the effectiveness of the proposed novel degradation model, we trained a deep blind ESRGAN super-resolver, and deployed it to conduct super-resolution on both synthetic images with diverse degradations and real low-resolution images. Experimental results demonstrate that the proposed new degradation model significantly enhances the practicality of deep super-resolvers, offering a robust alternative solution for real-world single image super-resolution (SISR) applications.

搜集汇总

背景与挑战

背景概述

RealSRSet是一个用于图像超分辨率研究的数据集，包含20张真实低分辨率图像，通过添加高斯噪声、JPEG压缩和模拟摄像机传感器噪声来合成退化图像。该数据集旨在验证新的退化模型，以提升深度盲超分解析器在实际应用中的性能，为真实世界超分辨率任务提供解决方案。