SPANet

从单幅图像中去除雨痕已经引起了相当大的关注，因为雨痕会严重降低图像质量并影响现有户外视觉任务的性能。虽然最近基于 CNN 的 derainers 报告了有希望的表现，但出于两个原因，deraining 仍然是一个悬而未决的问题。首先，现有的合成降雨数据集在模拟真实降雨特征（如雨形、方向和强度）方面的真实性有限。其次，没有对真实降雨图像进行定量比较的公共基准，这使得当前的评估不那么客观。核心挑战是不能同时捕获真实世界的雨/清洁图像对。在本文中，我们以两种方式解决单幅图像去雨问题。首先，我们提出了一种半自动方法，该方法结合了时间先验和人工监督，从每个输入的真实雨图像序列中生成高质量的干净图像。使用这种方法，我们构建了一个包含约 29.5K 雨/无雨图像对的大规模数据集，涵盖了广泛的自然雨景。其次，为了更好地覆盖真实雨纹的随机分布，我们提出了一种新颖的空间注意力网络（SPANet），以局部到全局的方式去除雨纹。大量实验表明，我们的网络在对抗最先进的去雨方法方面表现良好。

Removing rain streaks from a single image has attracted considerable attention, as rain streaks severely degrade image quality and impair the performance of existing outdoor vision tasks. Although recent CNN-based derainers have reported promising performance, deraining remains an open problem for two reasons. First, existing synthetic rain datasets have limited authenticity in simulating real-world rain characteristics such as the shape, direction and intensity of rain streaks. Second, there is no public benchmark for quantitative comparison on real rainy images, which makes current evaluations less objective. The core challenge lies in the inability to simultaneously capture real-world paired rain/clean image pairs. In this paper, we address the single-image deraining problem in two ways. First, we propose a semi-automatic method that combines temporal priors and human supervision to generate high-quality clean images from each input sequence of real rainy images. Using this method, we construct a large-scale dataset containing approximately 29.5K rain/clean image pairs covering a wide range of natural rain scenes. Second, to better cover the stochastic distribution of real rain streaks, we propose a novel spatial attention network (SPANet) to remove rain streaks in a local-to-global manner. Extensive experiments demonstrate that our network performs favorably against state-of-the-art deraining methods.

搜集汇总

背景与挑战

背景概述

SPANet是一个专注于单幅图像去雨任务的数据集，通过半自动方法构建了约29.5K个真实雨/无雨图像对，覆盖广泛的自然雨景，旨在提高去雨模型的真实性和评估客观性。该数据集还提出了一种新颖的空间注意力网络（SPANet），以局部到全局的方式去除雨纹，并在实验中表现出色。数据集由香港城市大学和大连理工大学于2019年发布，相关论文和资源公开可用。