The Research Of Single Image Dehazing Algorithm Based On Deep Learning

In haze weather,due to the existence of a large number of floating particles in the air,the light will interact with these suspended particles during the propagation process,causing the light to be scattered,and finally the scene light information reaching the imaging device will be damaged.Therefore,the captured images have problems such as low contrast,low definition,and loss of details,which affect the subsequent further processing and application of the images.Therefore,dehazing the hazy image enables it to be applied to high-level image processing tasks.In this paper,based on deep learning,a single image dehazing algorithm is studied.The main work is as follows:(1)Aiming at the inaccurate estimation of model parameters by the deep learning-based non-end-to-end image dehazing algorithm,and the problem of image spatial information retention in the current end-to-end image dehazing algorithm based on deep learning,this paper proposes a single image dehazing model based on high-resolution network,called De HRNet.De HRNet is divided into multiple branches,the resolutions between different branches are different,the branches with different resolutions are connected in parallel and multi-scale fusion is performed at the end of each stage.This paper adds a new stage to the original network to make it better for image dehazing.The newly added stage collects the feature map representations of all branches of the network by upsampling to enhance the high-resolution representations instead of only taking the feature maps of the high-resolution branches,which makes the restored haze-free images more natural.The experimental results show that De HRNet has obvious dehazing effect on hazy images.(2)Aiming at the end-to-end dehazing network model based on deep learning with too many parameters,and the problem of excessive or incomplete dehazing in the case of uneven fog concentration distribution,this paper based on the deep cross-scale fusion network proposes a lightweight image dehazing model,called Light-DCSFN.Light-DCSFN replaces the ordinary standard convolution in the original network with a depthwise separable convolution.And through the network to predict the fog density map of the input hazy image,instead of directly obtaining the haze-free image through the network to solve the problem of dehazing in the case of uneven fog concentration distribution.At the same time,to better extract the fog density feature and avoid the problem of color distortion in the restored haze-free image,a new color feature extraction module is added in front of the original network.First,the RGB three-channel image of the input hazy image is obtained by segmentation,and then input to the newly added module.It is used to extract the color shallow features between different channels.The experimental results show that Light-DCSFN not only has obvious dehazing effect on hazy images,but also has a short dehazing time for a single image.(3)Aiming at the problem that the image dehazing work is rarely combined with real life,and the dehazing effect of the Light-DCSFN dehazing model is further tested,this paper proposes a license plate that can be used in both sunny and foggy weather.This paper proposes a recognition system and a method for judging whether the image is hazy based on the dark channel prior knowledge.The system is divided into a module for judging whether the image has fog,an image dehazing module,a license plate positioning module,and a license plate character recognition module,which are used for license plate recognition in sunny and foggy weather.