基于增强多尺度生成对抗网络的单幅图像去雾

Enhaced Multi-scale Generative Adversarial Networks for Single Image Dehazing

雾容易导致采集的图像质量下降,并且包含的雾会影响后续的图像分析.单幅图像去雾是计算机视觉领域的经典问题之一.本文提出一种增强的多尺度生成对抗网络用于图像去雾.该方法不依赖物理散射模型,去雾网络由生成器、判别器和增强器3个部分组成.其中,增强器有助于采样多种特征使不同尺度的特征细节融入到结果,提升去雾图像在颜色和细节上的复原效果.增强器被分别嵌入到生成器和判别器,全局生成器和局部生成器融合生成一个由粗到细的高分辨率去雾图像,多尺度判别器用于监督生成图像.在真实世界和合成含雾图像数据集上的大量实验结果表明,提出的方法得到的去雾图像具有满意的主观视觉质量,并且利用最新的去雾定量评价指标,也具有好的客观图像质量.

Haze easily leads to the quality degradation of captured images,and the haze contained in hazy images has side effects on successive image analysis.Single image dehazing is one of the classical problems in the field of computer vision.In this work,we propose an enhanced multi-scale generative adversarial networks for single image dehazing.It does not rely on physical scattering model,and the network is composed of three components,namely generator,discriminator and enhancer.Among them,the enhancer helps to sample multiple features so that multi-scale features are fused into the resultant images,improving the restoration effects of color and details.The enhancer is embedded into the discriminator and the generator,respectively.The global and local generators are fused to generate a coarse-to-fine high-resolution dehazing image,and the multi-scale discriminator is used to supervise the generated image.Extensive experimental results on real-world and synthetic hazy image datasets show that the dehazed images obtained by the proposed approach have desirable subjective visual qualities,and also have good objective qualities in terms of the most recent metrics for image dehazing evaluation.