Haze removal from single images based on a luminance reference model

Optical transmission estimation is a key procedure for removing haze from certain outdoor images. In this paper, we propose a novel transmission estimation model called the luminance reference model. We theoretically prove the validity of this model, which demonstrates that the transmission can be computed accurately given a correct estimate of luminance reference. We also design an experiment exploiting large quantities of nature images to study the statistical property of the luminance reference. By statistics we find out that the distribution of luminance references in natural haze-free images is intensive. Therefore, to keep the simplicity and efficiency of the algorithm, the luminance reference is fixed as a constant for a single image. By using a scene-dependent estimate of the luminance reference, our method can be applied to different types of images. We further propose a guided approach to rapid and robust computation of transmission maps. Our experimental results show that the proposed method is computationally efficient, while producing comparable visual results to the existing state-of-the-art, but more complex, methods.