Image Dehazing Algorithms With Structure-preserving Filtering

As one of the most widely used information carrier,images have played a more and more important role in today's society.However,it is not always to obtain ideal images due to the imaging environments.Optical imaging system is sensitive to bad atmospheric conditions such as fog,mist,and haze.Images acquired under these conditions are suffered from degraded contrast and visibility,which seriously affects its usage for humans and many vision systems.Image dehazing technique aims at removing weather effects from images,and recovering a clear visibility and real color of the scene.So it is highly desired in earth observation,video surveillance,digital entertainment and other computer vision systems.It has attracted much more attention in recent years and shown great potential in many practical applications.This thesis focuses on the physical model based single image dehazing problem,which is a typical ill-posed problem since the composition factors of the haze image are unknown.Image filtering is a popular way to estimate the variables in the physical model.However,most of the existing techniques cannot restore scene structures well,which may cause undesired artifacts in the dehazed images.The structure-preserving filtering based image dehazing methods are developed in this thesis to improve the quality of the recovered image.The main contributions are summarized as follows: 1.Most of the existing image dehazing methods solve the problem based on the prior information,which may fail in different real-world images.This paper presents a haze removal method without using any prior.The transmission map of the hazy image is obtained via non-local structure similarity regularization.Specially,the simple dark channel is decomposed using non-local total variation regularization and the based layer is the transmission map.The nonlocal neighbors are searched on the minimal color component.So the similarity constraint can restrict the output transmission map have similar depth structures with the image.The proposed method has been tested on a variety of synthetic and real-world hazy images.The results demonstrate the effectiveness of our proposed image dehazing algorithm.2.The existing image dehazing methods usually blur edges in the estimated transmission which leads to halo effects in the dehazing results.Besides,most existing methods suffer from noise and artifacts amplification in dense haze region after dehazing.To address these challenges,we propose a transmission adaptive regularized recovery method for high quality single image dehazing.An initial transmission map is first obtained by a boundary constraint on the haze model.Then it is refined by applying a non-local total variation(NLTV)regularization to keep depth structures while smoothing excessive details.Noticing that the artifacts amplification effect depends on scene transmission,a transmission adaptive regularization based on NLTV is proposed to simultaneously suppress visual artifacts and preserve image details in the final dehazing result.An efficient alternating optimization algorithm is also proposed to solve the regularization model.Thorough experimental results demonstrate that the proposed method can effectively suppress visual artifacts for degraded hazy images,and yields high-quality results comparative to the state-of-the-art dehazing methods.3.The edge-preserving image filtering technique can remove noises and textures whilst retaining image structures.The existing global filters usually impose restrictions on local gradients in a regularization term to smooth details and textures.But it also affects image structures and will cause artifacts in the smoothed image.To this end,a new optimization framework is proposed in this paper for structure-aware image filtering via a non-local gradient sparsity(NLGS)constraint.The smoothing effect is obtained by diminishing intensity variation between the non-local neighbors.To distinguish textures from major structures,the filter first finds neighbors for each pixel by using a structure similarity measurement.Then the sparsity constraint is imposed on the simplified non-local gradients in a regularization model.Due to the NLGS constraint,the proposed method can avoid the impact on meaningful edges.Hence,it can better preserve image structures.The proposed filter is applied to refine the initial transmission map.The scene depth stuctures are predicted by a convolutional neural network.Then the similarity measurement is built on depth stuctures to find suitable non-local neighbors.This can ensure the precise transmission map for image dehazing.4.The visible bands of a multispectral remote sensing(RS)image contaminated by haze are also suffered from great noise,while the near-infrared(NIR)band is much less affected.This phenomenon is rarely considered in the multispectral RS image dehazing methods available.On the basis of the special attributes of the RS multispectral images,a structural information preserving dehazing algorithm is proposed.The visible bands are filtered under the guidance of the NIR image to remove noise and enhance the structures on them.The filter output is used to optimize the transmission map to make it smoother.Finally,the dehazing operation is carried out using the updated transmission and enhanced visible images.Experimental results on various satellite images show the fine visual quality of the dehazed image,and prove the effectiveness and superior performance of the proposed method.5.Thin haze removal is a challenging task since the estimation of haze component is easily affected by ground features.To solve the problem,this paper develops an effective haze removal method for a single visible remote sensing image.Firstly,haze is considered as an additive contamination and can be represented by a haze thickness map(HTM).A ground radiance suppressed HTM(GRS-HTM)is then proposed for a more precise estimation of haze distribution.The haze component for each band is calculated via GRS-HTM and can be removed to recover the clear image.Several visible satellite images with different resolutions were tested to validate the effectiveness of the proposed method.The evaluation results with qualitative and quantitative assessments demonstrate that the proposed method is superior to the traditional methods,and can recover a haze-free image with high quality.