Abstract:
Fog is a collection of water droplets or ice crystals drapped from the air at or near the earth surface.
It significantly reduces the visibility of outdoor scenes and hence causes road accidents, and flight
delays. The light coming from a scene towards the camera is attenuated by fog and is blended with
the atmospheric light. Although, many single image fog removal approaches have been proposed
in the past, their restoration performance is considerably low since they fail to consider the image specific cues. To feed that information, we propose a simple and robust defogging framework by
using image processing techniques. In the initial phase, rough depth map of a foggy image is
estimated based on the density of the fog in local image regions. In the proposed depth map estimation techique , a3x3 patch is used to obtain the dark channel of the foggy image and hence the intensity of obtained dark channel is considered as the density of the fog in that image. Then a
super-pixel based image segmentation technique has been used to identify the local depth image
regions based on the obtained dark channel. In the next step, scene-specific dark channel and
transmission are computed based on the segmented local image regions. Thereafter, pixel with
highest intensity is selected as atmospheric light and guided filter is used to get refined estimated
transmission map. As the final step, a histogram normalization and white balancing post processing technique is used to enhance the restoration and colour balancing. This study is
evaluated on FRIDA benchmark dataset, which has four different types of road foggy images. The
peak signal to noise ratio (PSNR), and the structural similarity index (SSIM) are used to measure
the performance of the proposed approach. High PSNR gives better quality of reconstructed image
and high SSIM index represents high similarity between ground truth and recovered image. The
experimental results show that the proposed defogging framework outperforms state-of-the-art
approaches.
