一种基于自适应大气光和加权引导滤波的夜间图像去雾算法

夜间图像去雾对于夜间场景下无人驾驶、交通安防等有重要的工程应用价值。针对暗通道先验算法在夜间雾天场景下的失效问题，提出一种基于自适应大气光和加权引导滤波的夜间图像去雾算法。该算法首先基于图像亮度和饱和度联合求取信道图，并将信道图作为引导图对原图像进行引导滤波得到大气光分布图，为解决暗通道先验在图像亮区域的失效问题，引入亮通道先验矫正亮区域的透射率，为优化亮、暗通道透射率的融合，建立一种基于分段伽马矫正的融合权值计算方法，用于亮区域透射率的权值计算，并利用该透射权值加权得到图像的初始透射率；然后利用加权聚合引导滤波代替引导滤波细化初始透射率，通过基于相似度为滤波中心像素的邻域像素赋予权值，并在滤波聚合阶段采用加权聚合代替均值聚合，解决引导滤波弱化细小纹理而引起的边缘模糊问题；最后将复原图像转换到HSV空间，对亮度分量V进行均衡化调整，并对均衡化前后的图像进行线性加权获得最终复原结果。实验结果表明，所提算法大气光分布图估值合理，可有效反映夜间多光源场景下的大气光分布情况，图像亮、暗区域透射率计算准确，复原图像去雾彻底、纹理清晰，与经典算法对比显示，复原结果的峰值信噪比、信息熵、平均梯度和方差的最大提升幅度分别为49.4%、18.3%、172.3%、115%，综合指标优于所对比的算法。

A Night Image Dehazing Algorithm Based on the Adaptive Atmospheric Light and the Weighted Guided Filter

Nighttime image dehazing has important engineering application value for unmanned driving and traffic security. In order to solve the problem that the failure of the dark channel prior algorithm in foggy night scenes, a night image dehazing algorithm based on adaptive atmospheric light and weighted guided filtering is proposed. Firstly, the atmospheric light distribution map is obtained by using the channel map based on the brightness and saturation of image to guide and filter the original image. Bright channel prior is introduced to correct the transmittance of bright region to solve the problem of dark channel failure in bright region. In the light and dark channel optimization and fusion stage, a weight fusion method based on piecewise gamma correction is used to calculate the weight of the bright region. Then in order to solve the problem of edge blurring caused by the weakening of fine texture in the guiding process, the weighted aggregation guided filtering is used instead of the guided filtering to refine the initial transmittance. In the process of the weighted aggregation guided filtering, the neighboring pixels of the filtered center pixel are assigned weights based on the similarity, and weighted polymerization is used instead of mean polymerization in the filtering polymerization stage. Finally, the restored image is converted to HSV space to equalize and adjust the brightness component, and the final restoration result is obtained by linearly weighted polymerizing of the images before and after the equalization. The experimental results show that the established algorithm has a reasonable estimate of the atmospheric light distribution map, which can effectively reflect the atmospheric light distribution in the nighttime multi-light source scene. The transmittance of the bright and dark areas of the image is calculated accurately, and the restored image is completely dehazed and has clear texture. The new algorithm shows the significant improvement percentages compared with previous algorithms with the maximum increases being 49.4% for the peak signal to noise ratio, 18.3% for the information entropy, 172.3% for the average gradient and 115% for the variance of the restoration results, respectively.