RSBlur

训练基于学习的去模糊方法需要大量模糊和清晰的图像对。不幸的是，现有的合成数据集不够现实，并且在其上训练的去模糊模型无法有效地处理真实的模糊图像。尽管最近提出了真实数据集，但它们提供了有限的场景和相机设置的多样性，并且为不同的设置捕获真实数据集仍然具有挑战性。本文介绍了RSBlur，这是一种具有真实模糊图像和相应的清晰图像序列的新颖数据集，可以详细分析真实模糊和合成模糊之间的差异。通过RSBlur数据集，我们分析了引入真实和合成模糊图像之间差异的各种因素，并提出了一种新颖的模糊合成管道，以合成更逼真的模糊。我们还表明，模糊合成中不同因素的影响以及我们的合成方法可以改善真实模糊图像的去模糊性能。

Training learning-based deblurring methods requires large quantities of paired blurred and sharp images. Unfortunately, existing synthetic datasets lack realism, and deblurring models trained on them fail to effectively handle real-world blurred images. Although real-world datasets have been proposed recently, they offer limited diversity in scenes and camera settings, and capturing real-world datasets for different settings remains challenging. This paper presents RSBlur, a novel dataset containing real blurred images and their corresponding sharp image sequences, which enables detailed analysis of the discrepancies between real-world and synthetic blurs. Using the RSBlur dataset, we analyze various factors that introduce discrepancies between real-world and synthetic blurred images, and propose a novel blur synthesis pipeline to generate more realistic blurs. We further demonstrate that the impacts of different factors in blur synthesis and our proposed synthesis method can improve the deblurring performance on real-world blurred images.

搜集汇总

背景与挑战

背景概述

RSBlur是一个专注于图像去模糊的真实数据集，提供大量真实模糊图像和对应的清晰图像序列，用于分析真实与合成模糊之间的差异。该数据集由Pohang University of Science and Technology于2022年发布，旨在通过新颖的模糊合成管道改进去模糊模型的性能，适用于计算机视觉和图像处理研究。