Varicolored Haze

有各种各样的图像除雾数据集，但很少有专门为去除杂色雾度而设计的数据集。为了解决这个问题，我们建立了一个大型的杂色雾度图像数据集，该数据集同时具有大规模的合成和现实世界的杂色雾度图像。在本文中，我们将静脉曲张的雾霾图像分为三类: 灰色浓雾，橙色烟雾和蓝色雾。我们通过Eq从驻留数据集中选择6000图像进行雾度模拟。1.对于每个干净的背景，我们生成具有随机散射系数的所有三个彩色雾度图像，从而获得18000的杂色雾度图像进行训练，并535图像进行测试。此外，为了评估真实世界的雾度图像，我们从互联网和现有数据集收集数据，以获取1254的静脉曲张真实世界的雾度图像以进行训练和评估。

There exist a wide range of image dehazing datasets, yet few are specifically designed for the removal of colored haze. To address this issue, we establish a large-scale colored haze image dataset that includes both large-scale synthetic and real-world colored haze images. In this paper, we classify colored haze images into three categories: gray dense fog, orange smog, and blue haze. We select 6,000 images from an existing dataset via Equation (1) for haze simulation. For each clean background, we generate all three types of colored haze images with random scattering coefficients, resulting in 18,000 colored haze images for training and 535 for testing. Furthermore, to evaluate real-world haze images, we collect data from the Internet and existing datasets to obtain 1,254 real-world colored haze images for training and evaluation.

搜集汇总

背景与挑战

背景概述

Varicolored Haze是一个专门用于去除杂色雾度的图像数据集，包含18000张合成的三类雾霾图像（灰色、橙色、蓝色）和1254张真实世界雾霾图像，由华中科技大学于2022年发布并配套CVPR2022论文。