Research On Rain And Haze Removal Method Of Single Image

In the age of artificial intelligence,driverless cars are a hot research topic for the automotive industry and can significantly improve the safety and efficiency of transport systems.In the field of unmanned driving,image information acquisition and processing is critical and directly affects the judgment of the unmanned system on road information.The quality of the image is therefore crucial for driverless cars,yet it is often greatly compromised by uncertainties such as weather(e.g.rain or haze).In recent years,with the rise of deep learning,research results on important theories and related techniques for image processing of intelligent transportation(e.g.motor vehicle windows,surveillance footage,etc.)have been rewarded with the ability to better remove rain or haze from images.However,the presence of raindrops on the windows on a rainy day is often accompanied by the effects of haze,which can obscure the background information of the image.Combined rain and haze removal algorithms for images have not been reported,so how to make the captured images unaffected by the combination of rain and haze is of great importance to the field of unmanned vehicles.In this thesis,based on the full investigation of theories related to image derain and image dehaze at home and abroad,the joint rain and haze removal problem is proposed for the image degradation caused by rain and haze weather,and related research is carried out,with the following main innovations.(1)A new single-image rain and haze model is proposed.Most of the current studies only consider the influence of raindrops or haze on the image,but in real life,car window images in rainy weather are not only affected by raindrops,but often accompanied by the interference of water vapour and haze in the air.To address the problem of image degradation caused by rain and haze,this thesis innovatively proposes a new single-image rain and haze model in the field of image processing at home and abroad,which no longer considers only the influence of a single rain or haze,but fully investigates the joint effect of multiple factors,and provides a theoretical basis for the single-image rain and haze removal algorithm proposed in this thesis.(2)A rain and haze removal algorithm for a single image is proposed.To address the fact that existing rain removal algorithms do not consider the effect of haze and that the traditional dark channel a priori theory is deficient for processing sky regions or dense haze of images,this thesis considers the optimization problem of joint rain and haze removal based on the rain and haze model we have built.The algorithm is framed by an integrated multi-tasking algorithm,which improves the taking of atmospheric light by combining fixed and training parameters,derives a new transmission map,and proposes an optimised visual attention network to solve the joint raindrop and haze removal problem.Experimental results show that the method in this thesis can effectively remove both rain and haze from a single image.