一种基于无监督学习的两阶段解耦图像去雾算法研究

A Two-Stage Disentanglement Image Dehazing Algorithm Based on Unsupervised Learning

利用卷积神经网络进行监督学习是解决图像去雾问题的一种常用的解决方法。然而,现有的方法大多主要使用成对的合成雾霾数据集,这可能不能准确地代表真实雾霾天气的场景。针对这一问题,本文提出了一种基于无监督学习的两阶段解耦去雾网络。该网络由三个子网络组成,它们分别将观测到的雾霾图像分解为无雾霾图像层、透射图层和大气光层。同时,该网络分为恢复无雾图像和透射图两个阶段。在第一阶段,利用嵌入的暗通道先验来获得无雾霾像和透射图的粗略估计。在第二阶段,通过两个子网络对第一阶段的结果进行细化,以产生更精确的无雾霾图像和透射图,而大气光则由另一个子网络直接估计。此外,本文还设计了一种新的多尺度注意力模块,作为细化无雾图像的子网。多尺度注意力模块在自注意力中执行多尺度的标记聚合,以捕获不同尺度的特征。实验结果表明,本文提出的网络获得了有效的雾性能和令人满意的视觉效果,且在PSNR、SSIM和主观视觉效果方面均优于现有的无监督去雾方法。

The use of convolutional neural networks for supervised learning is a commonly used method to solve the problem of image dehazing. However, most existing methods mainly use paired synthetic haze datasets, which may not accurately represent real-world haze weather scenarios. This article proposes a two-stage decoupling dehazing network based on unsupervised learning to address this issue. This network consists of three sub networks, which respectively decompose the observed haze images into a haze free image layer, a transmission layer, and an atmospheric light layer. Meanwhile, the network is divided into two stages: restoring fog free images and transmitting images. In the first stage, rough estimates of haze free images and transmission images are obtained using embedded dark channel priors. In the second stage, the results of the first stage are refined through two sub networks to produce more accurate haze free images and transmission maps, while atmospheric light is directly estimated by another sub network. In addition, this article also designs a new multi-scale attention module as a subnet for refining haze free images. The multi-scale attention module performs multi-scale label aggregation in self attention to capture features of different scales. The experimental results show that the proposed network achieves effective defogging performance and satisfactory visual effects, and is superior to existing unsupervised defogging methods in terms of PSNR, SSIM, and subjective visual effects.