Research On Haze Removal And Atmospheric Turbulence Mitigation

Outdoor optical imaging systems are inevitably affected by atmospheric scattering under complex weather conditions such as haze or fog,which will result in degradation like contrast decrease,dynamic range reduction,color distortion and blur,in the observed image or image sequence.Meanwhile,in long distance imaging process,observed image sequence often suffers from random geometric distortion and space-time-varying blur caused by atmospheric turbulence.Air pollution,haze and emission reduction is becoming a global issue,which is a long and arduous task.Obviously,it is especially urgent to explore technical approaches to reduce the influences on imaging quality caused by fog,haze,turbulence,etc.,which has been paid great attentions in machine vision and image processing in recent years.Research on the degradation mechanism and the restoration approaches of outdoor images under complex weather conditions has very important academic significance and extensive application potentials.It can provide not only clearer and more effective input data(image/image sequence)for existing machine vision and image processing applications,but also theoretical and technical support for the effective operation of all-weather optical imaging systems.Progress of dehazing research has been relatively slow because dehazing is a difficult problem.In 2009,the dark channel prior proposed by Chinese University of Hong Kong and Microsoft Research Asia won the best paper award of CVPR 2009.It greatly encouraged researchers in dehazing,but also lifted the starting point and increased the difficulty of the research.Under the above background and supported by China Scholarship Council,major international joint project of National Natural Science Foundation of China and Shaanxi Provincial research project,two key problems in outdoor scene observation: single image restoration in hazy days and scene reconstruction from atmospheric turbulence degraded image,are systematically and deeply investigated in this dissertation.The main research contents and contributions of this study are as follows:(1)In order to overcome the excessive enhancement problem of conventional histogram equalization technique,a contrast enhancement dehazing algorithm based on histogram optimization is proposed.By treating the process of contrast enhancement as the solution of a three-criteria optimization problem,the contrast of input hazy image is enhanced by generating a modified histogram that is close to the histogram equalization and adaptive gamma correction results,but also make the residual between histograms of input image and dehazed image small.Both subjective and objective evaluations of experimental results indicate that the proposed HOCE method is capable of avoiding over-enhancement in conventional histogram equalization algorithm,and can also effectively improve the contrast of the hazy image,while preserving color consistency of the scene.(2)The formation and physical model of hazy images is intensively studied,in order to overcome the limitation that existing priors and assumptions may be invalid in practical dehazing applications,a novel single hazy image restoration algorithm based on nonlocal total variation regularization is proposed.Atmospheric veil is estimated using nonlocal total variation based optimization,which is capable of overcome the excessive smoothing of image details caused by conventional local algorithms.By utilizing the similarity between image blocks,the nonlocal operation is able to eliminate staircase and halo artifacts in the dehazed images.It also ensures the smoothness of the atmospheric veil and better preserves edge and detail information in the image at the same time.After that,in order to deal with the limitation that existing priors may be invalid for bright regions of the scene,a combined boundary constraint is introduced for single image dehazing by taking both geometric and natural characteristics of outdoor images into consideration.According to imaging geometry,the geometric boundary of transmission map is obtained by estimating the relationship between scene object depth and the vertical distance of scene pixel and horizon.The combined boundary constraint is then constructed using the complementarity between the geometric and the natural boundary.Accurate transmission map is obtained by optimizing a nonlocal based regularization and haze-free image can thus be restored according to the estimated transmission map.Finally,in order to solve the noise magnification problem during scene radiance restoration,a transmission-aware image denoising method based on non-local optimization is proposed.Based on the assumption that noise distribution and transmission is correlated and image blocks with similar depth should have close color and brightness,non-local structure-aware regularization is introduced to suppress noise in the background region.Experimental results demonstrate that the proposed nonlocal total variation regularization based dehazing method is capable of suppressing halo artifacts while preserving details in the restored image.In addition,the proposed combined boundary constraint is capable of estimating a more accurate transmission map,compared with other state-of-art dehazing approaches.At last,the proposed transmission-aware image denoising method is able to significantly reduce the noise amplification in the background region.(3)Existing single scattering model has limitation on explain blurring effect in the images obtained in dense hazy days or from long distance,therefore,imaging process and degradation model based on multiple scattering is systematic studied and a novel hazy image restoration method based on multiple scattering is proposed.Blurring effect caused by atmospheric multiple scattering is usually described using atmospheric point spread function.In this dissertation,the kernel of atmospheric point spread function is approximated by generalized Gaussian distribution.Blur-free image is then generated using image deconvolution.Clear dehazed scene image is finally restored through hazy imaging model.Experimental results demonstrate that multiple scattering based dehazing method is capable of restoring a more clear and detailed scene radiance.(4)In order to reduce random non-rigid geometric distortion and time-space-varying blur caused by atmospheric turbulence in long distance imagery observation,a novel turbulence mitigation algorithm based on matrix decomposition and image block fusion is proposed.A Bspline based non-rigid image registration method is firstly applied as a preprocessing to reduce local geometric deformation in the observed imagery sequence.Secondly,the registered image sequence is decomposed into a low-rank scene matrix and a sparse turbulent matrix via matrix decomposition.A stable scene structure image and a turbulent image sequence are extracted from the corresponding matrices.Random noise in the turbulent image sequence is eliminated by averaging all frames.Based on the distortion position extracted from the averaged turbulent sequence,the sharpest turbulence patches in the registered frames are selected according to local clarity and similarity index.These selected patches are then enhanced with guided image filter.Finally,blur in the scene structure image is reduced using blind deconvolution,and restoration result is obtained by fusing the deblurred scene structure image and the enhanced image blocks together.Experiments indicate that the proposed approach is capable of eliminating geometric distortion and reducing scene blur caused by atmospheric turbulence degradation effectively.Compared with existing state-of-art approaches,the details and texture information in the scene are better restored by the proposed method.