一种基于改进非局部雾线先验的新型图像能见度检测算法

A Visibility Detection Algorithm Based on Non-Local HazeLines Priority

为了提高图像能见度检测的准确率,本文提出一种利用单幅日光条件图片检测能见度的方法,该方法基于非局部雾线先验和大气成像理论。首先通过改进非局部雾线先验选择图片中用于能见度估计的目标域(ROI);然后计算雾图像的透射率和恢复为清晰图像后的透射率,并计算二者的比值;最后,将清晰大气的消光系数与该比值相乘,得到被观测雾场景的大气消光系数,最终得到场景能见度。该方法一方面解决了传统方法需要摄像机参数标定或者大数据量训练的问题,另一方面相比采用暗通道先验的能见度估计方法,该方法在估计能见度时保留场景的全局性,并一定程度上避免了暗通道先验对光照的敏感,可以较好地在日光户外场景下使用。本算法平均准确率在89%以上,相较基于暗通道先验的比值法,平均准确率提升1%。

In order to improve the accuracy of image-based visibility estimation algorithm,a new method based on single image of daylight conditions is proposed.This method is based on non-local hazelines prior and atmospheric imaging theory.Firstly,the ROI in image is selected through non-local hazelines prior.Then,the transmittance of fog image and restored clear image are calculated,and the ratio of the two is calculated as well.Finally,the extinction coefficient of clear atmosphere is multiplied by this ratio to obtain the extinction coefficient of the observed fog scene.Eventually,scene visibility is obtained.This method requires neither camera parameter calibration nor training with large datasets.The proposed method retains scene globality in estimating visibility in comparison with the visibility estimation method using dark channel prior,and it also avoids the sensitivity of dark channel prior to illumination to a certain extent,so it can be better applied in outdoor daylight scenes.The average accuracy of this algorithm is over 89%,which is 1%higher than that of the ratio method based on dark channel prior.