Research On Natural Image Dehazing And Quality Assessment

Outdoor images often suffer from various distortions,such as low contrast,blurred details,dim color,which are caused by the widespread haze weather.Degraded images make it unable to meet the requirements of traffic surveillance,intelligent vehicle driving,aerospace and aviation.Research of natural image dehazing has been a hot issue because of its wide application prospects.In order to efficiently evaluate the performance of different algorithms and images on haze removal,research work of dehazed image quality assessment has a practical significance and practical value.In this thesis,we carry out our research based on problems that dehazed images often suffer from unclear details and incomplete boundary,and problems that existing dehazed image quality assessment algorithm cannot estimate the performance of oversaturated dehazed images.The author's major contributions are as follows:A novel image dehazing method based on weighted guided image filter with an improved guidance image,is proposed.To deal with unclear details and incomplete boundary in dehazed images that traditional dehazing algorithms based on the guided image filter encountered,we propose one scene depth optimization method based on weighted guided image filter with an improved guidance image.Firstly,the scene depth map,which suffers from serious block effects around boundary region and is not smooth enough in smoothing region,is roughly estimated based on the dark channel prior.Guidance image provides the foundation for edge remaining and texture smoothing in the method of haze removal based on guided image filter.Because of the prominent nature of RTV model in edge detecting and texture smoothing,we put it into optimizing the guidance image.The model that combined the proposed guidance image with the weighted guided image filter is used to optimize the rough scene depth map,which solves the problem that weighted guided image filter is not suitable for scene depth optimization.At last we can recover the haze-free scenes with the refined depth map using the atmospheric scatting model.The refined depth map by the proposed method is superior to others in texture smoothing and salient structures and edges remaining in complex region.The proposed is better to obtain haze-free scenes with high visibility,distinct details and natural visual effects.An image dehazing method based on relative total variation regularization optimization is proposed.To solve “halo” artifact that most guided image filter and regularization methodsbased on local operation suffer from,we propose one scene depth optimization method based on RTV model.Firstly,the scene depth map estimated from dark channel prior is compensated adaptively,to improve the underestimated depth map and noise influences processed by the small window.Texture of the compensated scene depth is so rich that it needs to be smoothing.As the relative total variance RTV model smooth texture and extract edges from the complex region based on a global scale,we put it into optimizing the compensated scene depth map,which efficiently alleviate the “halo” artifact caused by the local block operation.At last we recover haze-free scenes based on the atmospheric scatting model.The refined depth map is near-smooth in smooth region,and its structure is protruding.Haze-free scenes are also well with complete structures and rare “halo” artifact.A dehazed image quality assessment algorithm based on rank learning strategy is proposed.To solve problems that existing methods cannot obtain haze density accurately and estimate oversaturate distortion,we extract features from aspects of haze density and oversaturate distortion of dehazed images.Existing dehazed image quality assessment methods depend on accurate subject score,which results in inaccurate,unable and inconsistent predicted results.In order to solve these problems,dehazed image quality assessment is converted to classification and sorting issue between images based on the paired sorting learning idea.Firstly,according to proposed features,we estimate the characteristic vectors of paired images in proposed database,and corresponding labels that reflect quality of paired images.Then we learn the mapping relationship between image features and its classification result using random forest classification model.Finally,the final ranking results of images are determined through inverting paired classification comparison result in turn,which is predicted based on voting strategy.Experimental results fully validate the accuracy of features and database in this thesis,and are highly consistent with human visual perception by comparing the objective dehazed image quality assessment results of the proposed and traditional algorithm.