Image Dehazing Algorithm Based On Model Fusion

With the deterioration of air quality,haze weather has become more frequent in China's vast areas,and the problem of haze has become the main problem of atmospheric pollution in China.Under haze weather conditions,the suspended particles have the function of scattering and absorption in the haze,which makes the outdoor image show lower contrast and limited visibility,resulting in lower image quality collected on haze days.The quality of the collected image is easy to affect,and then it will limit the extraction of effective information from the image.Dehazing technology is necessary to enable computer systems to better identify and extract relevant image features.At present,the methods of image dehazing are mostly based on the atmospheric light scattering model,and the core of the model is to solve the atmospheric light and transmission map,so that the original haze-free image can be restored according to the model.However,remove haze technology is an ill-posed problem,which will cause the transmission map estimation based on some prior information to not be solved accurately.The biggest deficiency of these a prior are the incorrect estimation of transmission map and the common problem of transmission map saturation.In order to solve the above problems,this paper proposes a model fusion method for remove haze.Firstly,from the physical model of atmospheric light scattering formed by haze image,this paper analyzes the characteristics and differences of different atmospheric light calculation methods according to the characteristics of haze image,the common methods of atmospheric light calculation are sorted out and the optimal method is selected to estimate atmospheric light.Secondly,for the haze images at different haze levels,in order to obtain accurate transmission maps,this paper constructs two methods of model fusion for solving the problem of uneven haze.In the study of image haze removal based on model fusion: The first model is an accurate estimation of the transmission map by fusing the pixel mode model and the multilayer perceptron model.This method is based on the prior knowledge of the dark channel.By using a block of a predefined size to perform a rough transmission map estimation,the transmission map estimation in the block can be used to further optimize the transmission map at the pixel level using pre-calculated fast-guide filter values.The initial transmission map of the pixel mode model can be obtained.The depth resolution of the resulting initial transmission map is very low.Because it ignores the problems of neighboring pixels,therefore,the initial transmission map is estimated by multilayer perceptron model to obtain the final transmission map to improve the depth resolution,to solve the problem of transmission map mistake estimation.Finally,the contract extension strategy is used to enhance the contrast of the image.The second model combines block mode model and pixel mode model.Firstly,the input image is multi-band decomposed to extract the base layer.The base layer decomposed by the guided filter is used as the input image of the estimator,and the transmission map is estimated by block mode model and pixel mode model.The block method and the pixel method have their own advantages and disadvantages when estimating the atmospheric transmission map,respectively.The block mode causes the halo on the edge of the de-haze image,and the pixel mode always overestimates the actual haze degree,resulting in the phenomenon of oversaturation of the de-haze image.The haze-free image is restored by introducing a pair of weighted optimal fusion models based on the physical model of atmospheric light scattering.Finally,MATLAB tools were used for simulation.In order to verify the performance of the algorithm,this paper uses three standard test functions on three data sets to compare and analyze with other commonly used algorithms.The results of each algorithm and the corresponding tests data are obtained.The experimental results show that,in terms of image restoration quality and the robustness of the algorithm,subjective evaluation and objective analysis show that compared with other algorithms,the algorithm in this paper has the characteristics of high visualization performance and strong robustness.