Research On Image Defogging Algorithm Based On Deep Learning

Images taken in foggy environment will appear color attenuation,contrast reduction and other phenomena,which will affect the determination of the target,thus affecting all aspects of our daily life.Therefore,effective defogging has important theoretical and practical significance,as well as extensive application value and commercial value.The development of defogging technology can effectively improve the quality of images taken in foggy environment and improve the visibility of images,so as to better meet our daily life needs.The research content of this paper is as follows:(1)Aiming at the problem of poor generalization effect of synthetic haze image,an image de-fog algorithm based on improved domain adaptive is proposed.Due to the problem of domain migration,the model trained from the synthetic data can not generalize well to the actual fogged data.In this paper,an image defogging algorithm based on domain adaptive improvement is proposed.The algorithm consists of an image translation network and two image defogging networks.Firstly,an improved cyclegan network is applied to convert images from one domain to another to bridge the gap between the two domains.The translation module discriminator adopts the spectral normalized Markov discriminator(SN-Patch GAN)to improve the network generalization ability.Then the images before and after translation are used to train two deep image de-fogging networks with consistency constraints.At this stage,the features of clear images(such as dark channel prior and image gradient smoothing)are utilized to incorporate real fogged images into the de-fogging training to further improve the adaptability of the domain.By end-to-end training of image translation network and de-fogging network,better image translation and de-fogging effect can be obtained.Experimental results on synthetic and real images show that the model performs well.(2)Aiming at the problem of image detail loss,a defogging algorithm based on multiinput and multi-scale is proposed.The "coarse to fine" image de-fogging strategy has been widely applied to single image de-fogging network structure.However,the gradual upsampling from the bottom subnet to the top subnet inevitably leads to the loss of image detail.Therefore,this paper reexamines the strategy from "coarse to fine" and proposes a multi-input multi-scale image de-fogging algorithm(MIMS-Unet).The MIMS-Unet model has two distinct features.On the one hand,using multi-input multi-scale images for network input can significantly improve the overall model’s de-fogging performance,although it will increase the computation amount.On the other hand,a contextual background module is proposed to capture contextual information and recover more image details.Experimental results show that the model achieves good results in quantification and visualization,and avoids color distortion after fog removal effectively.