雾天退化模型参数估计与CUDA设计

Parameter Estimation of Fog Degradation Model and CUDA Design

针对基于物理模型的去雾方法大多采用统计或假设等先验信息获取模型参数精度较低的问题,提出一种非假设的雾天退化模型参数估计方法.为了尽可能准确地获取大气光值和透射率值,首先采用四叉树算法求解大气光值;随后利用预训练的卷积神经网络获取粗略透射率图,并使用引导滤波算法优化透射率图;最后通过大气散射模型逆向求解获取复原图像.实验结果表明,文中方法在去雾各项性能指标上表现均衡,不仅提高了雾天图像的清晰度和亮度,而且可以有效地避免"晕轮效应".算法时间性能实验表明,文中算法CPU效率比其他去雾算法提高40%+,应用CUDA并行设计将耗时的引导滤波算法并行化后效率有显著提升,处理分辨率大小为640×480(单位为像素)雾天图像仅需0.048 9 s,可直接迁移应用于视频去雾处理,满足视频处理的实时性要求.

Defogging algorithms based on atmospheric model always had atmospheric light and medium transmission limited by statistical or hypothetical information.Hence a non-hypothetical parameter estimation method was proposed.For precisely acquiring these parameters,the atmospheric light was solved by a quad-tree algorithm firstly.Secondly,a pre-trained convolutional neural network was proposed for estimating the transmission map optimized by the guided filtering algorithm further.Finally,by reversely solving the atmospheric scattering model,the de-fogging image was obtained.Experiments show that the proposed method has balanced performance on each index.It not only improves the degree of foggy image definition and brightness,but also efficiently avoids the Halo effect.Time performance also analysis indicates that,compared to other defogging algorithms,efficiency of our algorithm using CPU has increased 40%at least.After parallelizing the time-consuming guided filtering algorithm through CUDA,the efficiency has improved remarkably which can process a fog image with the resolution of 640×480 pixels only in 0.048 9 s.It can be directly applied to video processing to meet real-time requirement.