Image Dehazing Algorithm Based On Dark Channel Prior And Deep Learning

The introduction of the digital world,the huge growth of computer technology,and the introduction of artificial intelligence have all led to the widespread implementation of a variety of smart technologies,shown in every sector of our life.Machine vision systems have found use in a wide variety of industries,ranging from the relatively inconsequential to the essential aspects of daily life,such as image processing software,home security systems,and road monitoring systems,as well as in the manufacturing industry,aviation reconnaissance,military combat,and other fields.The outdoor computer vision system,on the other hand,scatters the light from the impurity particles in the air.This causes the images captured both inside and outside to appear gray and white,which affects the intelligent development of the vision system.Because of this,dehazing the hazy images obtained by the vision system on days when there is fog is of utmost importance,and it also has a significant impact in the practical world.Image dehazing has emerged as the primary focus of research in the field of machine vision systems as a means of enhancing the overall picture quality captured by these systems.The causes of haze as well as the effect that it has on photographs in hazy situations are investigated in depth in this research.In the beginning,the fundamental information and the physical model of air scattering theory are presented,and the imaging principle in a hazy environment is discussed in detail.The dark channel before dehazing algorithm as well as the deep learning image dehazing technique have both been examined in great detail at this point.This algorithm’s previous flaws have been resolved,and their solutions have been presented in deeper level.In the result,the dehazing effect of the upgraded algorithm as well as the effect of the older method are contrasted and examined using artificial subjective evaluation and objective evaluation indicators.The following is an outline of the primary research contents of this thesis:(1)An adaptive neighborhood dark channel and gradient-guided optimization of transmittance dehazing are two solutions that have been proposed in order to address the issue of the dark channel prior dehazing algorithm providing an inaccurate estimation of the atmospheric light values present in foggy images with large portions of the sky and bright regions.Firstly,adjust the size of the neighborhood window so that it is equal to 5% of the size of the original image that is being input pix 15pix×15pix of the neighborhood window ofΩ(x)in the original dark channel prior algorithm becomes an adaptive block,so as to obtain a more accurate dark channel value.Secondly,the value of atmospheric transmittance t(x)is computed with precision using an enhanced gradient guiding technique.The enhanced method adds an energy function to the guided filter and an edge maintenance term to the energy function in order to acquire more accurate transmittance and ensure that the dehazed image retains a large number of edge details.Finally,the transmittance of the large sky area and bright areas in the foggy image is corrected,and the accurate atmospheric light value is obtained using the quadtree decomposition algorithm,followed by the image being restored using the atmospheric scattering model.The experimental findings reveal that the enhanced method is fast in operation,and the color of the restored image’s sky area is more realistic,as well as the image’s clarity and contrast.(2)An enhanced convolutional neural network dehazing approach is suggested to overcome the difficulties of incomplete dehazing and erroneous transmittance computation in existing deep learning picture dehazing systems.The enhanced network model is divided into three stages: Initially,a shallow feature extraction will be carried out,as this is the first stage.In this stage,the input foggy image is given a vector convolution operation(Conv)that is carried out through three convolutional layers in order to generate the foggy image areas and various other properties of a shallow depth.The second stage perform a multi-scale convolution operation on the shallow feature map obtained in the first stage,and each layer will output 16 feature maps;then perform a maximum pooling operation on the feature maps that follow the multi-scale convolution layers in order to reduce the complexity of the network degree;the third stage is the fusion of shallow features and deep features,which is the most important stage in the neural network.Two feature extraction branches are added after the shallow network Conv1 and Conv2,and these two branches can obtain the image information output by Conv1 and Conv2 respectively.Additionally,a spatial pyramid pooling network is added to the branch to ensure that the input image size of the convolution network is consistent,and two more feature extraction branches are added after the shallow network Conv1 and Conv2 to obtain the image information output by Conv1 and Conv2 respectively.Fusion of deep and shallow features should be achieved.As the result,the revised neural network dehazing model avoids the issue of inadequate dehazing caused by a lack of picture information,makes the dehazed image more natural,and provides a great aesthetic appeal.