基于生成对抗网络的雾霾场景图像转换算法

本文提出了一种新的基于生成对抗网络的雾霾场景图像转换算法.生成对抗网络GAN作为无监督学习的方法,无法实现图像像素与像素之间映射,即生成图像不可控.因此,基于模型的加雾算法存在参数不确定性和应用场景局限性,本文提出了一种新方法的新应用,利用生成对抗网络实现图像转换.该方法基于生成对抗网络GAN模型,改进了GAN的生成器和判别器,进行有监督学习,以训练雾霾图像生成像素与像素之间的映射关系,实现无雾图像与有雾图像之间的转换.以图像加雾为例,本文分别设计了生成网络和判决网络,生成网络用于合成有雾图像,判决网络用于辨别合成的雾霾图像的真伪.考虑到雾霾场景图像转换的对应效果,设计了一种快捷链接沙漏形生成器网络结构,采用无雾图像作为生成网络输入,并输出合成后的有雾霾图像;具体来看,将生成网络分成编码和解码两部分,并通过相加对应间隔的卷积层来保留图像的底层纹理信息.为了更好地检验合成雾霾图像的真实程度,设计了漏斗形全域卷积判决器网络,将合成图像和目标图像分别通过判决器辨别真伪,采用全域卷积,利用神经网络进行多层下采样,最终实现分类判决,辨别图像风格.此外,本文提出了一种新的网络损失函数,通过计算GAN损失和绝对值损失之和,以训练得到更为优秀的图像转换结果.GAN损失函数的作用是使生成对抗网络GAN模型训练更加准确,而雾霾图像合成算法实际上是一个回归问题而非分类问题,生成器的作用不仅是训练判决器更加灵敏,更重要的是要生成与目标图像相似的图像.因此利用优化回归问题的绝对值损失函数,作用是为了准确学习像素间的映射关系,避免出现偏差和失真.最后本文对多类不同图像进行图像的雾霾场景转换并进行评估,分别测试该算法的图像加雾和去雾效果,并与其他算法进行对比测试.对于加雾效果,在合成场景、虚拟场景下,与软件合成效果进行对比,本文算法效果明显比软件合成效果好,不会出现色彩失真;在真实场景下,本文算法与真实拍摄的雾霾天气进行对比,结果十分相近;并且与其他GAN图像转换算法进行对比,本文算法具有明显的优势.同样本文算法在去雾效果上优势也十分明显.结果表明,本文所提基于生成对抗网络的雾霾场景图像转换算法,在主观效果和客观指标上均具有明显优势.

Image Conversion Algorithm for Haze Scene Based on Generative Adversarial Networks

In this paper,synthesis of hazy image based on generative adversarial net is described in detail.As an unsupervised learning method,the generative adversarial networks cannot learn the mapping between image pixels,which means the generation of images is uncontrollable.Therefore,because of the uncertainty of the parameters of the algorithms and the limitation of the application scenarios,a new application of the new method is proposed in this paper.The networks can learn the mapping between input and output image,and learn a loss function to train this mapping.The algorithm is based on GAN,and the improved generator and discriminator are proposed.And supervised learning is carried out to train the mapping between pixels in hazy images and haze-free images.Taking synthesis of hazy image as an example,we propose two networks:the one is generative network,which is used to generate hazy images,and the other one is discriminative network,which is used to identify the images.Here,considering the corresponding effect of image conversion,a hourglass-shaped fast link generator network is designed,which uses the fog-free image as the input.Specifically,the generated network is divided into two parts,encoding and decoding,and the underlying texture information of the image is preserved by adding corresponding convolution layers.Then,a funnel-shaped global convolutional judger network is designed to test the results of smog image synthesis.The composite image and the target image are respectively identified by the decider.The global convolutional neural network is used for multi-level downsampling to achieve classification and discerning image style.Meanwhile,the loss function of network is also modified.By calculating the sum of GAN loss and ABS loss,better results can be gotten.The loss function of GAN is to make the training of GAN model more accurate.However,the hazy image synthesis algorithm is actually a regression problem rather than a classification problem.And the generator is tasked not only to fool the discriminator but also to be near the ground truth Output.So in order to learn the mapping accurately and avoid deviation and distortion,the using of the ABS loss function is helpful.Finally,we analyze and contrast the experimental results carefully.This paper evaluates the transformation of different haze scenes,tests the effect of our hazing and dehazing algorithm,and compares it with other algorithms.For hazing effect,in synthetic scene and virtual scene,compared with the effect of software,the effect of our algorithm is obviously better than that of others,and there is no color distortion.In real scenes,the results of our algorithm are very similar to those of the real foggy and hazy scene.What’s more,compared with other GAN image conversion algorithms,our algorithm has obvious advantages.Similarly,it can be seen that the effect of our haze removal task is also very obvious.Experiments demonstrate that the proposed algorithm has better performance than state-of-the-art methods on both synthetic and real-world images qualitatively and quantitatively.