Image Dehazing Using Generative Adversarial Networks

In recent years,the hazy weather occurs frequently in China,and image dehazing gradually becomes a research hotspot.Hazy weather reduces the color saturation and contrast in images,and many image details are lost.In addition,such weather greatly affects the fields of satellite remote sensing monitoring,target recognition and tracking,and traffic monitoring.Therefore,image dehazing has very important practical application value and research significance.The image dehazing algorithm based on traditional methods can be divided into two categories:one is based on image enhancement,the other is based on physical model.The dehazing algorithms based on image enhancement enhance the hazy image directly,while the dehazing algorithms based on physical model produce the dehazed image through the prior information.The general applicability of these dehazing algorithms is relatively poor,and the dehazed image obtained by these algorithms has problems such as low brightness,halo and color distortion.The algorithm of image dehazing based on deep learning is to directly learn the mapping relationship between hazy image and clean image or between hazy image and transmission image,which avoids the traditional method of designing image features by hand,but the dehazing effect of this method on real hazy image is not stable.In this paper,the image dehazing algorithm based on the generative adversarial networks and the image dehazing algorithm based on the atmospheric scattering model are studied and improved.The main contributions are as follows:（1）In this paper,a multi-level image dehazing algorithm using a conditional generative adversarial network is proposed.Firstly,the hazy image generates the intermediate image K through the generator network,in which the intermediate image K is the joint estimation of the transmission image and the atmospheric light value,which can reduce the cumulative error caused by multiple intermediate parameters.Secondly,the intermediate image K obtains the dehazed image through the improved atmospheric scattering model,and carries on the adversarial training and the reconstruction restriction to the generator network and the joint discriminator network.In addition,Structural Similarity Index Measure is added to the loss function of the generator work to increase image details.（2）In this paper,the above algorithm is compared with several advanced image dehazing algorithms.The experimental results show that the image dehazing algorithm proposed in this paper achieves better dehazing effect in both the synthetic hazy image test set and the real hazy image test set.Specifically,for the synthetic hazy image test set and the synthetic hazy image test set,the dehazing algorithm proposed in this paper not only has a better subjective visual effect,but also is ahead of other advanced dehazing algorithms in the objective evaluation index.In addition,we have conducted contrast experiments on the network.When the whole network contains K+L_a+L_ssim+L1*,the dehazing effect is better.Finally,we verify the operation efficiency of the algorithm.It can be seen from the experimental results that the proposed algorithm is not only effective but also very efficient.