Research On Single Image Haze Removal Using Dark Channel Prior

More than 70% of the information obtained from the outside world comes from vision.With the rapid development of the industry and the improvement of the living standard,the frequency of foggy weather is accelerated and the degree is deepened.Because of the haze weather,the images obtained by the outdoor imaging systems appear the phenomenon of contrast dropped,the color shift and the details missing and so on.In order to be able to extract accurate and sufficient information from these images,it is necessary to dehaze the fog images so that it can be applied to traffic monitoring,remote sensing monitoring and so on.In this paper,the main work is based on the principle of dark channel proposed by Kaiming He.Firstly,we study the dehaze algorithm proposed by He deeply: the first step is to estimate the raw transmission.The second step is to obtain the refined transmission using soft matting and guided filtering.The third step is to estimate the atmospheric light.The fourth step is to obtain the clear image according to the model's parameters and the physical model.Then,we analyzed the disadvantages of He's two kinds of algorithms and two improvement schemes had been proposed:(1)estimate the atmospheric light using the quarter-weighted method so that the error caused by the existence of large light non-sky area in the image can be avoided;(2)estimating the refined transmission using Neighborhood Similarity Dark Channel Prior(NSDCP)so that the high complexity caused by soft matting can be reduced,and the halo phenomenon caused by over smoothing of guided filtering can be eliminated.Then,obtain the clear image using the atmospheric scattering model,the atmospheric light and the refined transmission those estimated by our algorithm.Finally,design the GUI interface to display the fog image,the image dehazed by the three algorithms,and the objective evaluation of the three algorithms.The objective evaluation index used in this paper is the running time of algorithm,peak signal to noise ratio,the average structural similarity index,the percentage of new visible edges,the normalized gradient mean of the visible edge,and the percentage of saturated black or white pixels.In this paper,the above three algorithms are used to dehaze the natural images,the images containing the buildings,the images containing the characters and the images of the vehicle license plates.The algorithm has the following advantages:(1)the time complexity is greatly reduced(an image of size 600*450,compared with He's two kinds of algorithms,the algorithm proposed by us decreased to 89.5% and 35%);(2)the halo phenomenon caused by over smoothing of guided filtering can be eliminated;(3)we can not only keep the quality of the dehazed image,but also obtain more detail information.At the same time,there is a need to improve the algorithm here:(1)the dehazed image may appear the phenomenon of color saturation,(2)the window size choose for estimating the refined transmission would affect the quality of dehazed images.