Research On Image Dehazing Based On Dark Channel Prior And Neural Network

As an objective reflection tool of the external world,images are widely used in modern life.It is of great significance to convert blurred images into clear images that can obtain more effective information by technical means.With the development of artificial intelligence,computer vision task intelligence is widely used in areas such as autonomous driving and video surveillance.However,due to the constraints of various factors in the shooting environment where the computer equipment is located,outdoor computer vision systems are easily affected by extreme weather such as haze.In recent years,haze weather has occurred frequently in my country.In the case of fog and haze weather,a large number of water molecules,PM2.5 and other impurity molecules in the air will affect the straight line propagation of light,resulting in blurred images,color distortion and contrast.Such problems as drop,loss of detailed information,etc.,will not only affect people’s visual perception of images,but also seriously affect the recognition and feature extraction of computer vision systems.In response to these problems,this paper studies the existing classical traditional dehazing algorithms and dehazing algorithms based on convolutional neural networks and improves the shortcomings of the existing algorithms.The proposed algorithm is designed based on the atmospheric scattering model and has obtained certain academic achievements.The specific work of this paper is as follows:(1)Aiming at the problems that the dark channel prior algorithm has insufficient transmittance estimation when dealing with large areas of the sky,and the restored image in haze weather has a halo effect where the depth of field is abrupt,the image details are not rich,and the restored image is dark as a whole.An image dehazing algorithm based on Lab space,dark channel prior theory,multi-scale Retinex and two-dimensional gamma function.First,the image is transferred to the Lab space,and the Canny operator is used to detect and extract the edge of the image to protect the edge information.In the Lab space,the multi-scale Retinex algorithm is used to process only the brightness component of the non-edge image.Secondly,the brightness component of the fusion edge information is obtained,the influence of the brightness component on the image is eliminated,and the image to be processed is obtained.Then,the rough estimated transmittance of the image to be processed is obtained through the dark channel prior theory,and the cross bilateral filter is used for optimization to obtain the accurate transmittance,which effectively suppresses the halo effect.The atmospheric light value is accurately obtained by using the improved quadtree subspace search method,which avoids the problem that the restored image is dark due to the overestimation of the atmospheric light value.Combined with atmospheric scattering model,a haze-free image is recovered.Finally,the brightness value is corrected in the HSV space by the two-dimensional gamma function,which eliminates the phenomenon of low brightness in the local area of the partially restored image.The experimental simulation shows that the proposed algorithm can effectively restore the detailed information of the image,suppress the halo effect,the color is natural and the image brightness is uniform,and the degree of dehazing of the restored image is relatively thorough.,which proves the effectiveness and feasibility of the algorithm.(2)In view of the fact that traditional algorithms rely too much on prior information,and the prior information is invalid when processing images obtained in some shooting environments,this paper proposes an end-to-end dehazing network model based on attention mechanism.First,the atmospheric scattering model is deformed,and the atmospheric light value and transmittance are combined into a whole to avoid the error interaction between parameters.Secondly,the feature information is efficiently extracted by two parallel multi-scale convolution kernels in the network model.Then,the residual structure is used to solve the problem of missing feature information,which improves the utilization of feature information and avoids the phenomenon of gradient disappearance.At the same time,an attention mechanism is added to learn the feature information in each channel,and different weights are assigned to different channels to optimize the feature information and enhance the effective feature information.Finally,the two sets of feature information are fused and the haze-free image is restored by using the atmospheric scattering model.The experimental results show that the proposed lightweight end-to-end dehazing network model can dehaze the degraded image,and the restored image has natural color,rich details and suitable saturation.The effectiveness of the proposed algorithm is shown in the standard comparison,and the efficiency of the proposed algorithm is proved by comparing the running time of each algorithm.