Research On Haze-degraded Image Restoration Based On Deep Learning

With the rapid development of computer vision technology,target detection and other intelligent visual information processing systems have put forward higher and higher requirements on the quality of input images.However,in hazy environment,the particles of impurities suspended in the air will scatter the light,resulting in the degradation of the quality of the captured image,showing problems such as low contrast,low detail recognition,and color distortion.Such foggy images not only affect the subjective visual perception of the human eye,but also have a serious impact on various intelligent image processing systems.Therefore,it is of great practical significance to study the defogging process of foggy images.Single image dehazing is a challenging ill-conditioned problem.The difficulty lies in the fog residue on the edge of the object and the details of the image are blurred.The priori-based enhancement/recovery method has the problems of narrow application range and unavoidable error accumulation.The existing deep learning-based methods often suffer from incomplete dehazing and inaccurate detail restoration.This article aims at the above problems,based on deep convolutional neural networks,using end-to-end learning methods,and by designing appropriate network structures and loss functions,the network is guided to learn the non-linear mapping between hazy and clear maps faster and better,which enables the network to solve the problems of fog residues and inaccurate detail reconstruction better.The innovative work of this article mainly includes the following two points:I.First,by analyzing the existing deficiencies of the existing traditional algorithms and deep learning algorithms in the four aspects of defogging ideas,data synthesis,network structure,and optimization methods,we proposed an end-to-end attention-based dehazing algorithm.The network is based on a coding-decoding framework.By directly learning the internal mapping of hazy images and clear images,it avoids the problem of error accumulation caused by algorithms based on physical models.The feature re-weight module is used to increase the attention of the network based on the importance of the characteristics of each channel to the dehazing task,and to suppress the characteristic channels that contribute less to the dehazing.At the same time,no extra spatial dimension is introduced,which maintains the network's efficiency.In addition,the use of skip connections and concat operations avoids the problems that the existing algorithms have too little neuron receptive field and the gradient disappears when the network is too deep.The results of subjective and objective comparisons with various representative existing algorithms show that the image restored by our method is more complete in dehazing,the colors are more real and natural,and the visual effect is better.II.For the existing dehazing algorithms based on convolutional neural networks,haze is left on the edges of the object and the texture details are not clear.We proposes a new depthassisted estimation dehazing network.First,the degraded image is passed through a separately designed depth estimation network.which extracts the depth-related information of the image,and then inputs this information together with the hazy image to the dehazing network.Depth information can assist the dehazing network to learn the mapping between before and after image degradation,making the network easier to learn and better image restoration.At the same time,a designed edge loss is added to the loss function to enhance the extraction of detailed features,avoiding the unclear details of the existing algorithms.The residual module and multi-level and multi-scale pooling used in the network further enhance the ability of nonlinear fitting and neuronal receptive field.Experimental results show that the depth-assisted estimation dehazing algorithm proposed in this paper is superior to other algorithms in the evaluation of objective indicators of synthetic haze images and the evaluation of subjective restoration effects of real haze images.