Single Image Haze Removal Using Genetic Algorithm

Abstract

Single image based haze removal is a very challenging field due to the scarcity of ground truth images of the hazy images for the quantitative validation of the work in this field. A haze removal algorithm based on single image combined with the genetic learning model has been addressed in this thesis work. The basics behind this work is simply finding or deriving a depth map model that directly relates to the saturation and brightness features of the haze. The parameters of the model are learned with powerful evolutionary genetic algorithm which is a supervised learning method. Once the depth map is obtained from the given hazy image, the transmission of the scene can be easily restored, and the scene radiation representing the original image can be restored using the atmospheric model. In the atmospheric model, the scattering coefficient, which is the measure of extinction due to scattering of monochromatic radiation as it propagates through a medium that consists of scattering particles, plays a vital role in the dehazing performance. After the completion of the GP process, nearly optimal solution for depth model is obtained. The effect of varying scattering coefficient () on the dehazed images was also studied using 10 test images in the research. The results show that the visual quality of the dehazed image n by rate of visible edges, gradient ratio and SSIM is satisfactory in the range to 1.2 in terms of SSIM, toin terms rate of visible edges and 1.2 to 1.6 in terms of gradient ratio. The SSIM using the GP learning method was obtained 0.74986, whereas the CAP method resulted in SSIM of 0.547907. MSE of the dehazed images was found to be 0.024571 for the GP learning method, and for CAP method, it was found to be 0.050529. The quantitative results prove that the quality of dehazed images that are degraded by haze is improved by the proposed method based on GP learning strategy.