Analysis And Research On Transmission Model For Hazy Images

Outdoor images are usually degraded by the airborne particles (fog, smoke, etc.) in atmosphere. This degradation can greatly lower down the performance of video surveillance and image collection systems. Researchers have devoted great efforts to haze removal in order to improve the reliability and robustness of these systems. On the other hand, the presence of haze is a momentous part for images to keep depth with aerial perspective. Researchers are also interested in fog/haze simulation that can impose realistic effects to virtual scenes. Fog/haze simulation can generate Chinese brush, water color painting effects, and etc. on real world images.However, it is a pity that few researches can handle the work of haze removal and haze simulation simultaneously and the two fields have the same essence, the calculation of transmission. Therefore, in this paper, we propose the transmission model for hazy images, to remove and simulate haze. First, we reformulate the fundamental atmosphere scattering model, introduce the definition of max visibility and make connection between hazy images. Then we propose new natural transmission model (haze filtering model) and take color correction and sky compensation as assist to refine results of algorithms.For the transmission estimation, the kernel part of haze filtering, we employ two kinds of technical routes:one method simply uses dark channel prior to estimate transmission for high efficiency, furthermore, the fast algorithm of dark channel prior and guided filtering are employed to improve the efficiency. The operation is easy and fast which can reach 10-1s and meet the requirements of real-time; the other method adopts Gaussian Process Regression, which takes hazy images and transmission maps as the training sets and transmission as the output vectors. Furthermore, it applies multi-scale and multi-dimension feature vectors to learn the transmission map, and this method has higher applicability and generality. Experimental results show that our haze removal results present good clarity, fidelity and contrast, meanwhile, the haze simulated results look quite authentic and natural.