Image dehazing is a pre-processing step in computer vision tasks, that has attracted considerable attention from the research community. Existing CNN-based methods ignore haze-related priors and rarely use a coarse-to-fine scheme in a feed-forward architecture to remove haze due to increasing network depth and parameters. This results in sub-optimal dehazing results. To address these problems, a multi-scale attentive recurrent network is proposed for image dehazing, which consists of a haze attention map predicted network and a recurrent encoder-decoder network. First, by assuming that haze in an image is formed by multiple layers with different depths, the haze attention map predicted network is designed for generating the map with multiple stages via a multi-scale recurrent framework. Second, the haze attention map is viewed as the haze-related prior and guides the subsequent recurrent encoder-decoder network to be aware of haze concentration information. Finally, for leveraging the intermediate information and optimizing the dehazing result with less parameters and more robust features, the recurrent residual operations which pass the features of selected layers at the current time step to the corresponding layers at the next time step are applied in the recurrent encoder-decoder network for removing haze following a coarse-to-fine strategy. Experiments on synthetic and real images demonstrate that our method outperforms state-of-the-art methods considering both visual and quantitative evaluations. In addition, our method is also suitable for real-time processing.

图像去雾是计算机视觉任务中的一个预处理步骤，引起了研究界的广泛关注。由于网络深度和参数增加，现有的基于 CNN 的方法忽略了与雾霾相关的先验，并且很少在前馈架构中使用从粗到细的方案来去除雾霾。这导致次优的去雾结果。为了解决这些问题，提出了一种用于图像去雾的多尺度注意力循环网络，它由雾度注意力图预测网络和循环编码器-解码器网络组成。首先，假设图像中的雾霾是由具有不同深度的多层形成的，雾霾注意力图预测网络旨在通过多尺度循环框架生成具有多个阶段的地图。第二，雾霾注意力图被视为与雾霾相关的先验，并指导随后的循环编码器-解码器网络了解雾霾浓度信息。最后，为了利用中间信息并以更少的参数和更健壮的特征优化去雾结果，将当前时间步的选定层的特征传递到下一时间步的相应层的循环残差操作应用于循环用于按照从粗到细的策略去除雾霾的编码器-解码器网络。在合成和真实图像上的实验表明，我们的方法在视觉和定量评估方面都优于最先进的方法。此外，我们的方法也适用于实时处理。