基于暗原色先验的图像快速去雾

Fast image haze removal base on dark channel prior

目的针对暗原色先验去雾算法出现的边缘残雾、天空色彩失真以及速度较慢问题,提出一种快速有效的图像去雾算法。方法舍弃传统分块的思想,采用逐像素处理的方法估计透射率,并对估计值过低的透射率进行适当的增强。大气光采用效率更高的四叉树算法来求解。结果有效地解决了边缘残雾和天空色彩失真问题,相比其他算法,去雾后的视觉效果有所提升。透射率和大气光的求解速度都得到一定程度的提高,去雾速度是暗原色先验去雾算法的近4倍。结论实验结果表明,本文算法在保证良好去雾效果的前提下能大幅提升去雾的效率,节省去雾所花费的时间。对于大部分有雾图像,本文算法都能够达到较好的去雾效果,但在处理具有较大景深的图像时,远景部分的去雾效果欠佳。鉴于速度上的优势,本文算法适用于对实时性要求比较高的去雾场合。

Objective Significant research has been conducted in the field of image haze removal locally and internationally. However, haze removal methods, which have good effects, often take a long time. Moreover, effects of fast methods are general to reach the requirements of applications in many occasions. This study aims to present a method that has fast processing speed and improved effect based on mainstream haze removal methods. The haze removal method via dark channel that is previously suggested by He is simple and effective. However, it keeps residual haze near depth edges after haze removal. Moreover, it leads to color distortion in sky area and large white areas, which do not previously meet dark channel. In addition, its processing speed is slow. This study aims to solve these problems and to present a fast and effective method based on the method suggested by He. Method The appearance of residual haze is due to block thought adoption via the method of He and the assumption that the transmission keeps unchanged in a local patch. We can abandon block thought, cancel the minimum filtering operation, and use per-pixel processing method to estimate transmission map. The appearance of color distortion is due to extremely low estimation by the method of He for the transmission in the sky area and large white areas, which leads to subtle differences among pixel RGB color channels in these regions, which are magnified nearly 10 times. Thus, we can increase transmission of these regions properly. Atmospheric light can be estimated with quadtree algorithm, which is efficient. Result Our method has solved the problems of residual haze and color distortion effectively. Moreover, haze removal speed is enhanced greatly because operations of minimum filtering and soft matting or guided fihering in the process of estimating the transmission map are abandoned and the atmospheric light is improved by solving efficiency. The speed of our method is about four times of that of the method by He. Conclusion Experimental results show that our algorithm can greatly improve efficiency of haze removal and can save the time spend by haze removal under premise of good effect. Our method can keep good effect of haze removal for most haze images. However, if an image has deep scene depth, our method have a general manifestation on its distance scene. On account of the advantage in speed, our method is suitable for real-time demand higher occasions.