Single Image Dehazing Based On Variational Regularization Method

Image dehazing is an important research issue in the field of computer vision.Its main purpose is to reduce or even eliminate the adverse effects of haze on image quality.The existing dehazing algorithms can be classified into two kinds:image enhancement-based dehazing algorithms and physical model-based dehazing algorithms.Among them,due to the strong pertinence and superior recovery performance of physical model-based dehazing algorithms,it has gradually become the mainstream method in the field of image dehazing.However,there are still several problems to be optimized in these physical model-based dehazing algorithms,such as poor edge-perserving,weak artifact suppression and easily fail in large sky regions.These deficiencies are largely caused by the inaccurate estimation of transmission,so this thesis focuses on the accurate estimation of transmission.The main works of this thesis are as follows:1.Since many dark channel prior-based dehazing algorithms are poor in edge-preserving and artifact suppression,this thesis proposes a Total Variation?TV?modelTV-L^1,2 that fuses multiple regularization constraints to optimize the transmission.Due to the effect of multiple regularization constraints,thisTV-L^1,2model can not only preserve image details effectively,but also suppress the block effects and artifacts well.In addition,to overcome the problem that the dehazing algorithm is easily distorted in large sky regions,this thesis further introduces a tolerance mechanism into the hazy image formation model to improve the recovery results.2.To effectively improve the dehazing performance in large sky regions,this thesis further proposes an improved dehazing algorithm based on the soft segmentation and Joint Total Variation?JTV?model.This improved dehazing algorithm first uses soft segmentation method to refine the dark channel prior-based initial transmission,which effectively overcomes the problem of color distortion and unnaturalness in large sky regions.Furthermore,by introducing a constraint on the latent haze-free image and transmission,this thesis proposes an improvedJTV-L^1,2model based on theTV-L^1,2 model to jointly optimize the latent haze-free image and transmission.This model can well suppress the deviation caused by soft segmentation method while optimizing the transmission.3.Since the non-smooth optimization models are difficult to solve,the corresponding alternating direction method of multipliers-based numerical solution algorithm is proposed to efficiently solve theTV-L^1,2 model andJTV-L^1,2model.The numerical solution algorithm can decompose the complex non-smooth optimization problem into two easily solved subproblems,and solving these subproblems alternately.Then,the stable numerical solution can be obtained effectively while guaranteeing the convergence of the numerical solution algorithm.Numerous quantitative and qualitative experiments show that compared with several classical physical model-based dehazing algorithms,the proposed image dehazing algorithms can recover excellent haze-free images under different scenes.