Research On Deep Learning Models And Algorithms For Image And Video Dehazing

With the popularization of digital imaging equipment,the application of digital images and videos has penetrated in various fields,but images and videos taken in outdoor scenes will be affected by severe weather such as haze,which seriously reduces the visual quality of images and videos.The main task of image and video dehazing is to recover clear images and videos from haze images and videos.As only the haze image and video are known,image and video dehazing is an ill-posed and challenging problem.Although image and video dehazing algorithms have been proposed and achieved considerable progress,there are still some challenges that need to be improved,such as the problem of error propagation caused by separately estimating the transmission map and atmospheric light in the deep dehazing method based on physical models,the balance problem between network parameters and generalization,the generalization performance of the network model based on supervised learning when processing real haze images is insufficient,the problem of alignment of adjacent frames with the reference frame in video dehazing.To solve those problems,the main research work of this thesis is as follows:(1)An image dehazing model based on a conditional generative adversarial network is proposed to avoid the propagation of intermediate variable error.This method first uses the encoding and decoding network as the generation network model,and introduces a residual learning mechanism in the encoding and decoding features so that the generation network can better learn the deep features that help dehazing.Secondly,in order to reduce the number of parameters of the generation network model while ensuring its dehazing performance,a multi-scale haze image is used to guide the training of the decoding network to cut out the coding network.Finally,in order to take into account the visual and objective quality of the dehazing image,a gradient prior based on the L1 constraint and the VGG feature constraint are proposed.The training process of the proposed network adopts an end-to-end training method.A large number of experimental results show that the proposed method can effectively remove the haze in the image and generate high-quality clear images.(2)An image dehazing method based on a multi-scale recursive network is proposed,which balances the network parameters and model generalization.This method gradually removes the haze in the image by repeatedly using the same network structure on different resolution scales and does not increase the number of network parameters while improving the dehazing effect.In order to better train the multiscale recursive network,an auxiliary loss function is developed to guide the training of the multi-scale recursive network and improve the dehazing performance of the model.In addition,the smooth L1 loss and the perceptual loss are used to jointly optimize the training process.Experiments prove that the proposed method uses a small number of network parameters to achieve state-of-the-art performance.(3)Two methods of combining deep neural network and physical model of image dehazing are proposed to improve the generalization of the network model.The task-oriented network embeds the physical model of image dehazing into the deep neural network model in the feature space;the semi-supervised image dehazing method based on the physical model enables the deep neural network to train on the synthetic dataset in a supervised learning manner and train on the real dataset in a self-supervised manner according to the atmospheric scattering model.In addition,the task-oriented network is a hybrid network that embeds the spatially variant recurrent neural network between the encoding and decoding networks,which improves the generalization of the network model to restore the haze image.The semi-supervised image dehazing method improves the generalization of the dehazing network in processing real haze images and reduces the domain gap between synthetic datasets and real datasets through the joint training of the dehazing network,the transmission map estimation network,and the atmospheric light estimation network on the synthetic data set and the real data set.Experiments show that the two methods can effectively remove the haze from real images.(4)A multi-stage deep video dehazing model is proposed to solve the problems of temporal and spatial coherence of adjacent frames and video dehazing.This method first uses a two-stage fusion network to solve the problems of adjacent frame alignment and reference frame dehazing.To improve the quality of the dehazing video,the third-stage refinement network is used to further remove the haze in the reference frame.At the same time,to ensure the compactness of the network,the three fusion networks in the first stage adopt the mechanism of network sharing parameters.Finally,the perceptual loss and the smoothed L1 loss are combined to constrain the training process of the multi-stage network.Experiments show that the multi-stage video dehazing method can effectively remove the haze in the video and deal with the temporal and spatial coherence of adjacent frames.