Algorithm Research On Single Image Dehaze

With the development of science and technology, especially the development of industrialization, particles in the air escalating which resulst in fog and haze occuring frequently. Images captured in fog weather appear decreased contrast, color saturation, sharpness and other issues, which greatly reduced the value of image use. How effectively and quickly remove the affect on image caused by fog became a hot topic of computer vision. This cross-disciplinary front topics has broad application prospects.Firstly, this thsis research the physical processes of image degradation and mechanisms of foggy blur. Analyse the atmospheric scattering model, including the attenuation model and the airlight model, and on this basis, obtains the fog imaging model. Then, based on the atmospheric scattering model, the characteristics of fog image such as blur, color distortion and the decrease of contrast are analyzed. Next comes the content of dark channel prior and dehaze algorithm based on this prior. And we discuss the case of dark channel prior inapplicable, since dark colors is a prior of statistical laws, and it does not work well under some scenarios.For limits of dark channel prior, we propose a method based on HSI space. By a formula derived relationship between the thickness of the fog and reduction degree of saturation, and the proposed to use transmission values to enhance the saturation. In view of the situation that the brightness of dehaze image is lower than original image, the brightness enhancement algorithm is proposed based on the relationship between the reducing quantity of luminance and the transmittance. For the case of dark channel prior inapplicable, this thsis proposes split and separate treatment area of inapplicable, and then use image fusion to synthesize two partial results. To deal with the halo phenomena in some complicated scenarios, proposed to obtain dark channel on downsampled image. It proved that this method can effectively avoid the halo phenomenon caused by the size of filter window is too large.Finally, combining with subjective evaluation and objective evaluation, this thsis compares the results of dark channel prior algorithm and the proposed algorithm. Compare with dark channel prior algorithm, our algorithm can avoid the color noise generated when the dark color channels prior inapplicable, and also to avoid color shift problem. Compared with guiede filter, algorithm proposed in this thsis can avoid the halo phenomenon more effectively.