Adaptive Image Dehazing with Dark Channel Prior and Edge Components

Abstract

This paper presents an image enhancement technique to remove haze contained in an outdoor image based on the depth information estimated from dark channel and edge components. Dark channel prior (DCP) refers to a statistical observation that the pixels of a non-sky image patch in a haze-free outdoor image tend to show very low intensity in at least one of three color channels. Many existing DCP-based image dehazing methods attempt to estimate a transmission map, rather than the depth from the camera to the objects in the scene, which is optimized with a soft matting function to remove haze. The resulting dehazed images often suffer from halo artifacts due to depth discontinuity between near and far objects in the scene. The haze-removal effect on far objects can also be limited. The proposed image dehazing method estimates the depth information using the amount of haze measured by the DCP and the edge components of the objects since near objects are likely less affected by haze and therefore reveal stronger edge information. The estimated depth discontinuity is used to adjust the soft matting function to obtain more accurate transmission map and therefore enhanced dehazing effect. Experiment results show that the proposed dehazing method is effective to retain more image details preserved in the dehazed image with no halo artifact.