Proc. Computer Vision & Pattern Recognition Vol. 1, pp. 325-332 (2001). 1 Instant Dehazing of Images Using Polarization Yoav Y. Schechner, Srinivasa G. Narasimhan and Shree K. Nayar Department of Computer Science, Columbia University, New York, NY 10027 {yoav,srinivas,nayar}@cs.columbia.edu Abstract We present an approach to easily remove the effects of haze from images. It is based on the fact that usually airlight scattered by atmospheric particles is partially po- larized. Polarization filtering alone cannot remove the haze effects, except in restricted situations. Our method, how- ever, works under a wide range of atmospheric and viewing conditions. We analyze the image formation process, tak- ing into account polarization effects of atmospheric scat- tering. We then invert the process to enable the removal of haze from images. The method can be used with as few as two images taken through a polarizer at different ori- entations. This method works instantly, without relying on changes of weather conditions. We present experimental results of complete dehazing in far from ideal conditions for polarization filtering. We obtain a great improvement of scene contrast and correction of color. As a by product, the method also yields a range (depth) map of the scene, and information about properties of the atmospheric particles. 1 Introduction Recently there has been a growing interest in the analysis of images of scenes affected by weather phenomena. The main objective has been to enhance [6, 10, 15, 24] images taken in poor visibility, and even restore the clear-day vis- ibility of the scene [11, 12, 14]. It has also been observed that degradation of images by atmospheric scattering can actually be exploited to obtain information about the scene, particularly its range map [4, 12, 14]. Some image enhance- ment methods proposed in the past require prior information about the scene (e.g., distances [15, 24]). Other methods are based on specialized radiation sources and detection hard- ware [16, 23]. Computer vision methods have restored clear-day vis- ibility of scenes using neither special radiation sources nor external knowledge about the scene structure or aerosols [11, 14]. These methods rely only on the acquired images, but require weather conditions to change between image acquisitions. This can take too long to make dehaz- ing practical. They also require that the scattering properties will not vary with wavelength. This paper describes an ap- proach that does not need the weather conditions to change, and can thus be applied instantly. Moreover, in this ap- proach the scattering properties may vary with wavelength. Our approach is based on analyzing images taken through a polarizer. Polarization filtering has long been used in photography through haze [20]. Relying only on optical filtering is, however, restrictive: it is sufficient only on clear days, with weak light scattering (mainly due to air molecules), when the sun is ≈ 90 o to the viewing direc- tion [9, 20]. In these situations photographers set the po- larization filter at an orientation that best improves image contrast. In general, however, polarization filtering alone cannot remove the haze from images. Here, we obtain much more than optics alone can yield by analyzing the polariza- tion filtered images. The analysis of polarization filtered images has proved to be useful for computer vision. For example, it was used to analyze specularities [13, 17, 25], separate transparent and semi-reflected scenes [7, 18, 19], classify materials [26], and segmenting scenes [1]. We note that advances in po- larimetric cameras [1, 21, 25, 26] enable acquisition of po- larization information in real time. In this paper we model the image formation process, tak- ing into account polarization effects of atmospheric scat- tering in haze. We then use this model to recover the de- hazed scene, and also obtain information about scene struc- ture and atmospheric properties. Our approach does not re- quire modeling of the scattering particles’ size or their pre- cise scattering mechanisms. The principle is very simple: the image is composed of two unknown components - the scene radiance in the absence of haze, and airlight (the am- bient light scattered towards the viewer). To recover these two unknowns we need two independent images. We easily obtain these images because usually airlight is partially po- larized. The method only requires that the airlight induces some detectable partial polarization. We demonstrate re- moval of haze effects from a real scene in a situation where pure optical filtering (without applying our algorithm) does not suffice at all. 2 Theoretical Background 2.1 Airlight Polarization One of the causes for image degradation associated with atmospheric scattering is airlight. In this process, light coming from the illumination sources (e.g., the sun) is scat- tered towards the viewer [14]. Consider Fig. 1. The airlight