Abstract

In rainy conditions, especially at night with low illumination, the visual of images obtained by outdoor computer vision systems is degraded significantly, leading to a significant negative effect on the work of the outdoor computer vision system. In this paper, we develop a new rainy image model to describe rain scenes at night with low illumination. From this model, we propose a joint deep neural network-based method for single nighttime rainy image enhancement. First, a decom-net based on Retinex theory is employed for image decomposition, and the purpose of this sub-net is to extract the reflection image and the illumination image from the input image. Then, an enhancement net is proposed for illumination adjustment. The goal of this sub-net is to remove the negative effect (low visual) caused by low illumination. Finally, a symmetric sub-net termed multi-stream network-based contextual autoencoder is developed, where rain features are directly learned from the enhanced nighttime rainy images in a recurrent way. The goal of this sub-net is to effectively remove rain streaks from the illumination-enhanced image. The experimental results show the advantage and effectiveness of the proposed method, and evident improvements over existing state-of-the-art methods are obtained with the proposed method.

中文翻译：

---

基于联合深度神经网络的夜间夜间单图像增强方法

摘要

在下雨天，特别是在夜间低照度的情况下，由室外计算机视觉系统获得的图像的视觉质量显着降低，从而导致对室外计算机视觉系统的工作产生严重的负面影响。在本文中，我们开发了一个新的雨天图像模型来描述夜间低照度下的雨天。从该模型，我们提出了一种基于联合深度神经网络的夜间夜间雨天图像增强方法。首先，基于Retinex理论的分解网用于图像分解，该子网的目的是从输入图像中提取反射图像和照明图像。然后，提出了一种用于照明调节的增强网。该子网的目的是消除低照度引起的负面影响（低视觉效果）。最后，一个称为多流网络的基于上下文的上下文自动编码器的对称子网得以开发，在该子网中，以循环的方式直接从增强的夜间阴雨图像中学习雨状特征。该子网的目标是从照明增强的图像中有效消除雨水条纹。实验结果表明了该方法的优点和有效性，并且与现有技术相比，已有明显的改进。