Research On Multi-scale Depth Neural Network Defogging Algorithm Based On Residual Dense Block Group

In severe haze weather,haze particles in the air diffuse and absorb light,reduce the visibility and saturation of the image taken,severely distort the color,and seriously affect the quality of the image taken.Therefore,image fog removal has become a hot topic in the field of computer vision.Traditional fog removal methods use prior knowledge to make statistics on the characteristics of the image,so as to calculate a fog-free image,but the image will have serious color and detail distortion.In recent years,with the rapid development of in-depth learning,image fog removal technology has made significant breakthroughs and progress,and has overcome the time-consuming and labeling feature processing problem.However,there are still many drawbacks,such as the quality of the defogging image is not stable,and the model design is mostly based on atmospheric scattering model.When the parameters of atmospheric scattering model are not accurately estimated,the calculation of the neural network makes the error continue to be enlarged and the effect of the recovered foggless image is worse.In addition,most algorithms do not get good results in high fog concentration scenarios,and do not recover enough details and colors.In order to solve the above problems,this paper analyses the research results of in-depth learning fog removal methods at home and abroad,and makes the following contributions:(1)A multiscale feature fusion fog removal network based on dense residual groups is presented.De-fogging,especially at high fog concentrations,while maintaining image details is a more difficult task.We believe that an important reason for this is that global fog removal and local detail preservation are contradictory requirements.In order to preserve local details while removing fog,a multiscale method is used to extract three different scales of image features by convolution and downsampling.Based on the dense residual structure,a new Residual Dense Block Group(RDBG)is proposed for deep feature extraction at various scales to support short and medium term continuous memory in multiple residual dense blocks.Then these multiscale features are fused to reconstruct the defrosted image.The experimental results show that this algorithm performs well in image defrosting and is superior to the existing mainstream algorithms.(2)A two-stage fine fog removal network based on channel attention mechanism is presented.Although dense groups of residuals can extract enough features,the defrosted image still produces problems of inadequate shading and detail recovery in some bright areas.These problems are still caused by high frequency estimation errors.Therefore,on the basis of dense multiscale residuals,channel attention mechanism with emphasis on medium and high frequencies is added to improve the detail recovery of the output image.In addition,in order to make more efficient use of the medium and high frequency information in foggy images,a two-stage fog removal network structure is proposed: the output from the multiscale attention residual dense group network(the first-stage network)is channel joined with the original fog map,and as the input of the second-level network,the features are extracted and integrated again through the residual dense group and convolution layer,and the first-order network is obtained.The output of the network is modified slightly to further refine the fog removal and maintain the image details.The experimental results show that the network can effectively correct the details of the output image of the first level network,give the network the ability to process important features and details,and obtain better results in both subjective judgment and image index judgment.