Single nighttime image dehazing based on image decomposition

Abstract

Dehazing plays an important role in promoting the performance of outdoor computer vision systems. However, existing dehazing methods are targeted to daytime haze scenes, and are not able to improve visual effects for nighttime hazy images due to the unpredictable factors at night. In this paper, an effective single image dehazing framework based on image decomposition is presented for nighttime hazy images. First, the input single nighttime image is separated into the glow-shaped image and the glow-free nighttime hazy image using its relative smoothness constraint. Then, a novel structure-texture-noise decomposition model based on the exponentiated mean local variance is devised to split the nighttime hazy image into a structure layer, a texture layer and a noise layer, in which the structure layer and the texture layer are dehazed based on the maximum reflectance prior and the dark channel prior and enhanced in the gradient domain respectively. Finally, the dehazed structure layer and the enhanced texture layer are fused to produce a dehazed result. Experiments demonstrate that the proposed approach outperforms several state-of-the-art dehazing techniques for nighttime hazy scenes, especially in terms of noise suppression. Besides, the proposed algorithm is also capable of handling daytime hazy images and low-light degraded images.