Haze Removal Using Generative Adversarial Network

The problem of haze removal has been addressed in many computer vision researches. Haze removal is the process of eliminating the degradation present in hazy images and getting the clearer counterpart. The presence of haze distorts the image, and as a result, it will be difficult to apply various image processing techniques on such images. The challenging aspect in haze removal arises due to the lack of depth information in images degraded by haze. The earlier methods for haze removal include various hand-designed priors, usage of the atmospheric scattering model or estimation of the transmission map of the image. The limitation with these models is that they are heavily dependent on the assumption of a good prior. In recent years, various models have been proposed which effectively remove the degradation caused by haze in images using various convolutional neural network architectures. This paper reviews a model which performs haze removal on a single image using generative adversarial network (GAN). The main advantage of this method is that it does not require the transmission map of the image to be explicitly calculated. The model was evaluated using NYU depth dataset and 0-Haze dataset. The model was able to significantly enhance the quality of the images by generating the corresponding haze-free counterpart. The model was evaluated using the peak signal-to-noise ratio and structural similarity index.