Research Of Image Restoration Algorithm For Night Hazy Image

Images of outdoor scene those taken in hazy weather condition are degraded due to absorption and scattering of atmospheric aerosol particles. The degradation includes reduction of contrast and distortion of color, etc. At present, the research of image restoration algorithm for hazy image with uniform illumination (daytime) has already achieved significant results. However, the research of image restoration algorithm for hazy image with non-uniform illumination (nighttime) has little results. Nighttime haze removal is more difficult, since nighttime hazy images contain degradation more serious than daytime's due to illumination, resulting in lower brightness, lower contrast, uneven illumination, chromatic aberration, and higher noise, etc. Therefore, in this paper we study the image restoration algorithm for nighttime hazy image. This paper includes the follow two contents:(1) Depending on imaging law of nighttime hazy image, we give a new image degradation model for nighttime hazy image and accordingly present a novel dehazing algorithm based on illumination estimation. On the one hand, for the random distribution of multiple light sources at night, we present estimation of uneven illumination and color of light sources. Firstly, a multi-layer threshold segmentation method is proposed to detect the location of light sources. Secondly, rough light map is estimated according to attenuating law of light propagation and geometric assumption. Thirdly, brightness constraint, lower bound constraint and local smooth constraint is used to correct rough light map. Finally, the color of illumination is estimated by statistics of light source regions. On the other hand, for various assumptions of day dehazing algorithms don't apply to night image, an entropy-based scheme for nighttime images is proposed to obtain rough transmission map and the guided filter is used to correct rough transmission map. On the basis of estimation of uneven illumination and transmission map, we use the night haze imaging model directly to restore the clear image.(2)When use imaging model directly (model don't consider noise) to restore image, noise will be amplified. This is the problem of haze image restoration algorithms at present, especially the algorithms for night image with high noise. In order to solve the problem, we present a novel image restoration algorithm based on prior of image patch and EPLL optimization framework, dehaze and restrain noise at the same time. Firstly, for clear image library of general scene and similar scene respectively, the expectation maximization algorithm is used to obtain gaussian mixture prior based on image patch. Secondly, cost function is designed by two-norm and prior model and the recovery result is obtained by half quadratic splitting algorithm. Finally, a noise level estimation algorithm based on image quality evaluation algorithm is proposed to set parameters of half quadratic splitting algorithm for improve the visual quality of result.