Research And Implementation Of Image Defogging Technology Based On Conditional Generation Confrontation Network

With the rapid development of computer vision technology,it has been widely used in video monitoring,intelligent transportation,target tracking and other scientific research fields.However,outdoor computer vision studies are highly susceptible to weather conditions.Frequent haze weather in recent years is a very big threat to this research.In haze weather,the image captured by the camera device will have a severe degradation reaction.This is because the sunlight will been scattered heavily by small particles or small water droplets floating in the air.Therefore,in order to ensure the development of computer vision technology and improve the quality of hazy images,dehazing technology is very important.Based on the study of how haze forms and what the influence of haze,we proposed a new dehazing algorithm.The research content of this paper is as follows:(1)Based on the advantages of deep learning,we proposed an image dehazing algorithm based on conditional generative adversarial nets(CGAN)to learn the mapping between hazy pixels and haze-free pixels.This kind of mapping can fundamentally solve the influence of haze on the image.In order to accelerate the convergence of the network and improve the quality of dehazing results,we fuses different loss functions as the final loss functions of CGAN's generator.And in the process of fusing different loss functions,we use grid search method and random search method to determine the most appropriate proportional parameters of each loss function.(2)In order to improve the quality of dehazing results,a residual learning dehazing algorithm is proposed based on the atmospheric scattering model.This algorithm learns to infer the loss information which caused by haze from the hazy image.And after we get the loss information which called residual information too,we can add it linearly with the hazy image to get the final haze-free image.According to the residual learning,it is faster to learn residual information than to directly learn the mapping from hazy pixels to haze-free pixels.Furthermore,in order to prevent gradient explosion and gradient disappearance and ensure the convergence of the network,we add gradient clipping measures into the training process.(3)Based on the dehazing algorithm proposed in this paper,an image dehazing system is designed and implemented.The system can run the above two kinds of image dehazing algorithms and give a satisfactory dehazing result graph.What's more,the system can also score the dehazing results,to help users judge the results of dehazing.