PRIOR BASED APPROACH FOR SINGLE IMAGE DEHAZING

Abstract

In the past few year, with the increase in the living standards, the understanding of security is becoming important. Video surveillance became the most crucial in terms of security, it can be used in warehouses, malls, outdoor and traffic. But in bad weather conditions, fog has a large impact on the images of the video which can degrade the quality of image and hence disturb the work of security. Hence working on removal of fog and getting a clear image became one of the most important task for the researchers. This degradation can lead to a low-quality image, which also servers input to many computer vision algorithms. Hence, dehazing is very important step for an image. The degradation can majorly be seen using the transmission map, which is one of the crucial parameters of Dehazing using a single image. The estimation of the transmission map is an underlying issue, and lots of different prior are proposed for that. Among them, one of the widely recognized prior is DCP. DCP is very efficient, effective, and easy to implement algorithm but on the same side it has some limitation. In this project we will discuss limitation of DCP and explain multiple solutions given by researchers to overcome that limitations. After that we will discuss our proposed solution and the implementation of that. We gave a simple and efficient way of dehazing a single image based on color attenuation prior and the notion of a dark channel. We first find the ambient light using decision image and dark channel then computes the transmission using the color attenuation prior (CAP). CAP uses a depth map to compute the transmission. In the later stage, we refine the transmission using the bilateral filtering and restore scene radiance using the refined transmission map and ambient light. To evaluate our approach we had used two different parameters namely discrete entropy and FADE and done comparison using iv two other state of the art approaches. Results shows that our model outperform the other state of the art algorithms.