Research On Image Defogging Based On Random Forest And Kernel Regression

Under the conditions of fog and haze,the scattering of atmospheric particles will lead to reduced image contrast,poor visual effects,some important information is obscured by fog and cannot be identified.Therefore,the technology of fog removal has become a hot topic in the field of image processing and computer vision.Most of the algorithms cannot learn the parameters from the data,and the poor fogging quality and the long running time,a new method based on random forest and kernel regression is proposed in this paper.Firstly,the random forest regression model is constructed to predict the transmissivity of fog images;secondly,the atmospheric light is synthesized by vector method;finally,the clear image is restored according to the atmospheric illumination model.The main innovation work includes:(1)In view of the edge blurring and the block phenomena existing in the existing image denoising algorithms,a method of transmittance smoothing and thinning based on Gauss kernel regression is proposed.Firstly,according to the priori knowledge of the dark channel and the atmospheric illumination model,the rough estimate of the transmission is obtained.Secondly,in order to solve the problem of block phenomenon in the rough estimation,the Gauss kernel regression is used to deal with the problem.Finally,the transmittance is taken as the marker of the sample image.(2)In view of the existing image defogging method for lack of data driving,to learn the model parameters from the haze image,causing the defogging effect is poor,the long time,put forward the method of defogging haze image based on random forest.Firstly,the transmission parameters of the training samples are labeled by kernel regression smoothing;Secondly,the characteristics of the haze of the training samples are extracted,and the sample images are put back into the random sampling to construct multiple decision trees to form the random forest.Finally,for an input image,the average value of all decision trees is the result of random forest model.(3)In light of the uneven brightness distribution of fog and haze images,the light vector method is proposed to estimate the atmospheric light parameters.By estimating the direction and intensity of the atmospheric light,the parameters of the atmospheric light vector are synthesized.By comparing the defogging algorithm of qualitative and quantitative and several kinds of the most advanced,that image dehazing method random forest and kernel regression combination,improved fuzzy edge and long operation time,the problem of uneven brightness.