Unsupervised Attention Guided Image Dehazing Research

Outdoor scenes usually have turbid media remaining in the atmosphere,and haze is generated due to the absorption and scattering of the atmosphere.The image acquisition equipment is affected by the haze,resulting in reduced image clarity and contrast,and in severe cases,color and detail information may change or be distorted.In recent years,computer vision technology has been popularized into the application of various industries,these systems have become an indispensable part in roads,aviation,Unmanned aerial vehicles and other fields.Haze weather has seriously threatened traffic safety.To enable all kinds of computer vision systems to work normally in the haze environment,and make the subsequent image segmentation,target detection,image recognition and other algorithms run effectively,it is of very important practical significance and wide application value for image dehazing.Traditional image dehazing methods are mostly based on atmospheric scattering models or prior knowledge,but these are not always applicable,and model parameters cannot be accurately estimated,resulting in color distortion and poor results in dehazing results.Deep learning algorithms enable the network to automatically learn the parameters of the model or the task of image dehazing.And there are many unsupervised dehazing methods based on generative adversarial network,which can generate clear images in an end-to-end way.However,the previous dehazing network did not take into account the different density characteristics of different regions in real haze images,and only performed the overall image transformation,resulting in uneven dehazing,poor details,and color cast problems.In view of the above problems,this thesis proposes a new algorithm of dehazing.The innovation and contribution of this thesis can be summarized into three aspects:(1)This thesis explores the connection between attention mechanism and dehazing,and proposes an attention mechanism based on dark channels.In this thesis,the spatial attention mechanism used in image local conversion is applied to image dehazing,and the feasibility of attention mechanism in image dehazing is explored through experiments.Inspired by the dark channel prior algorithm,this thesis innovatively proposes an enhanced attention mechanism to focus on the haze area,and solves the limitations of the original attention mechanism in image dehazing.This thesis verifies the effectiveness of the dark channel attention method through experiments.It can accurately and quickly mark the concentration and area of haze,which makes it possible to dehaze the network partition components.(2)This thesis constructs two attention-based dehazing networks.Based on the architecture of the original cycle generation adversarial network,the generator and discriminator network structure is improved for the goal of image dehazing,and combined with the proposed dark channel attention mechanism.Relying on cyclic generation process,this thesis constructs two dehazing network based on the dark channel attention mechanism.The proposed networks avoid the problem of image color distortion and poor detail reduction caused by the overall conversion of the image by the traditional generation adversarial networks,and can pay attention to the different degree of change required for different haze concentrations.The proposed network structure can retain more detailed features,restore color information and improve the dehazing effect.Through a variety of comparative experiments,the thesis verifies the effectiveness of the proposed dehazing network structure.(3)Exploring the robustness of dehazing multiple types of images.This thesis uses the actual dehazing application scenario as the background,and according to the feature of image acquisition equipment that collects all-weather.This thesis divides the input images in the actual dehazing operation into three categories,and verifies the robustness of the proposed method and the comparison method for the three types of image conversion.Compared with other algorithms,the network structure in this thesis has significant improvements and enhancements in various scenarios.It can effectively convert hazy images,misty images,and haze-free images,and get the expected conversion effect,which is of great significance to the hybrid dehazing task of complex scenes in practical applications.