Single Image Dehazing Using Neural Network

The need for Single Image Dehazing came up as a result of hazy input images captured during foggy or hazy weather. This occurs due to the fact that certain dust particles and smog can easily scatter light, especially during morning haze, some firework or at the dawn time. Therefore, a hazy image gets piled over the original image. And hence, it becomes a challenging task to retrieve the original image from the input hazy image. Generally for single image dehazing, a massive dataset of input hazy image is required, the reason being Deep Learning is the backbone of the entire functionality of this concept. Deep Neural Networks require multiple hidden layers between the input hazy image and the output layer. Though Single Image Dehazing employs methods like polarization, prior based approach, extra information method, prior based method, learning based method have shown the greatest level of accuracy in recovering a clear image. Amongst the existing methods, polarization method and contrast based methods weren’t applicable in real time scenarios. Although, Dark Channel Prior based method was one of the most successful amongst the prior based strategies, it’s drawback was that it overestimates the thickness of the haze. In this paper, the main focus will be at comparing different Deep Learning methods, stressing upon various Convolutional Neural Networks, thereby giving a deep insight of various CNN strategies for retrieving the original dehazed image.