Single image dehazing is a complex and ill-posed task. Hazing is attributed to dust, fog and other environmental factors which severely degrades the images. Image hazing is a function of depth, where the visual contrast reduces rapidly as the depth of objects in the images increases. From a perspective of vision, this severely impacts feature retrieval based tasks. Degraded photos often lack visual appeal and offer poor visibility of scene contents. Thus, dehazing images forms an important task in consumer photography and a crucial preprocessing step in vision tasks. There has not been a lot of work in image dehazing in the domain of deep learning. Most of the deep learning approaches that exist for single image dehazing are based on multi-scale CNNs.
We explore this problem from the perspective of Generative Adversarial Networks for this first time and experiment with different formulations of loss functions and stacking GANs.