Research On Single Image Defogging Algorithm Based On Deep Learnin

With climate change and the rapid development of the world economy,the frequency of haze weather is gradually increasing,which makes the images acquired by image acquisition equipment appear fuzzy features,reduced contrast and color distortion.The single image dehazing task aims to restore clear images from hazy images with relevant algorithms.This task can be used as the front module of various high-level computer vision tasks.The existing methods based on deep learning overcome the shortcomings of traditional methods based on a priori knowledge design features by virtue of their strong feature learning ability,and further improve the image dehazing effect.However,there are two main problems in the existing methods.First,due to the limitation of model design,there are many network parameters and the training is difficult;Second,due to the lack of basic theoretical support,the model is prone to color distortion,haze residue and other phenomena for the restoration of real scenes,especially foggy images under heavy haze.In view of the above problems,this paper designs corresponding single image dehazing algorithms,the main contents are as follows:(1)Aiming at the problems of many network parameters and difficult training in existing methods,a progressive image dehazing network is proposed.By designing the basic dehaze unit,the number of basic dehaze units is controlled,and the iterative execution is carried out to achieve progressive dehazing.During the iterative execution of the basic dehaze unit,network parameters are shared,thus reducing the difficulty of network training.In the design of the basic dehaze unit,the Long Short-Term Memory network is introduced to transmit the features of different progressive stages,and the hierarchical features of the image are extracted by designing the encoding-decoding network.The atmospheric scattering model shows that image depth is very important for restoring clear images.In order to better restore the depth of the image,the depth map of the original hazy image is introduced to optimize the network.The proposed network does not depend on the atmospheric scattering model and outputs clear images in an end-to-end manner.Experiments show that the progressive image dehazing network can achieve effective dehazing on both synthetic datasets and real scene datasets,and is superior to the existing methods in subjective and objective evaluation.(2)For the limitations of existing methods in real scenes,especially in heavy haze,a two-stage image dehazing network based on Retinex theory is proposed.The proposed network consists of residual illumination map estimation and image dehazing.In order to establish a theoretical model of image dehazing,an image dehazing model based on Retinex theory is established by analyzing the relationship between Retinex theory and haze imaging,and a multi-scale residual illumination map estimation module is designed to obtain the image after preliminary dehazing.Then,a fine dehazing module is designed to optimize the rough dehazing image to obtain a clear image with more thorough dehazing and richer details.The proposed network does not depend on the atmospheric scattering model and outputs clear images in an end-to-end manner.Experiments show that the twostage image dehazing network based on Retinex theory has good generalization ability on different datasets,especially for heavy haze in real scenes,it can achieve effective dehazing.