基于B-样条曲线暗通道的图像去雾算法

针对暗原色去雾算法在高亮背景图像中出现光晕（Halo）效应的问题，提出一种基于B-样条曲线加gamma（γ）修正透射率去除Halo效应的算法。通过分析暗原色透射率图像的统计直方图，得出Halo效应的产生是由于HE算法透射率直方图中低灰度值占比较大，因此需要对直方图低灰度包络进行修正。寻找直方图中第1个低灰度波峰，在2个波谷间进行子区间分割，并以区间边界点作为样条区间节点计算三次样条曲线，以拟合低灰度包络。基于透射率图中修正位置对修正结果的关键性作用，采用改进的大津法从横向和纵向两个维度对透射率图进行分割，以确定修正区域，实现精准补偿，提高抗噪性能。根据不同雾图选择合适的γ修正系数修正拟合的灰度包络并对图像进行还原，验证不同修正系数对结果的影响。实验结果表明，当修正系数γ>1.5时，Halo效应能被有效减缓或消除，与HE算法相比，该算法的对比度、信息熵最大可分别提高10.1%和9.9%，算法并行度可达到M或N级。

Image Defogging Algorithm Based on B-Spline Curve Dark Channel

A modified algorithm for transmitting Halo removal based on the B-spline curve and gamma (γ) correction is proposed to solve the Halo effect of dark primary color defogging algorithm in bright background images.First, the statistical histogram of the dark primary color transmission image is plotted.It is known from the data that the halo effect is induced by the extremely high proportion of low gray transmission, so the low gray envelope of the histogram required modification.Hence, the first gray peak in the histogram is searched, and subinterval segmentation between the two valleys is conducted.The cubic spline curve is determined with the interval boundary points as spline interval nodes to fit the low gray envelope.Second, the position of correction in the transmission map plays a significant role in the results, and it is necessary to separate the area to be corrected from the transmission map.The improved Otsu method is then used to segment the transmission map from the horizontal and vertical dimensions to determine the correction area and improve the antinoise performance.Third, appropriate correction coefficients are selected based on different fog maps to modify the gray envelope and restore the image to verify the effects of different parameters.The experimental results show that the Halo effect can be suppressed or eliminated when the correction coefficient γ>1.5.Compared with the HE algorithm, the contrast and information entropy of this algorithm can be improved by 10.1% and 9.9%, respectively, and the algorithm parallelism can reach the M or N level.