基于交通场景区域增强的单幅图像去雾方法

针对当前已有的去雾算法在雾天道路图像的处理上易造成近处路面区域和远处天空区域亮度过低、处理程度偏强，而中远处区域去雾程度较低、亮度过高等问题，以基于深度学习去雾算法为基础提出一种结合雾天道路图像场景深度和道路图像特点的去雾算法。首先基于深度学习的去雾算法原理，构建卷积神经网络求取场景透射率；然后基于大气散射模型和透射率估计出图像深度图，且构造两个参数：上阈值和下阈值来将深度图分为中、远、近三个区域；再基于深度图的不同区域构造增强函数，来确定图像处理的增强幅度照；最后在传统的大气散射模型基础上结合增强幅度照来调节不同区域的复原强度得到优化后的处理图像。实验结果表明，所提算法可以在保证良好去雾效果的前提下增强道路图像的中远处区域，有效解决了去雾后雾天道路图像近处路面和远处天空的色彩失真、对比度过低问题，提升复原图像的视觉效果，并且与暗原色先验算法、均匀与非均匀雾的视觉增强算法以及典型的基于深度学习去雾算法相比具有更好的图像清晰化效果。

Single image dehazing algorithm based on traffic scene region enhancement

For the current dehazing algorithm easily results in low brightness of near road area and distant sky area with strong dehazing, and high brightness of middle and distant area with weak dehazing, based on a depth learning dehazing algorithm, a dehazing algorithm combined with image scene depth and road image characteristics of fog and sky roads was proposed. Firstly, based on the principle of dehazing algorithm of deep learning, a convolution neural network was constructed to calculate the scene transmittance. And then the image depth map was estimated based on the transmittance and atmospheric scattering model. Two parameters were constructed, the upper threshold and the lower threshold, to divide the depth map into middle, far, and near areas. Based on the enhancement function constructed by the different parts of the depth map, the enhancement amplitude of image processing was determined. Finally, based on the traditional atmospheric scattering model, the intensified illumination intensity was used to adjust the recovery intensity of different areas to obtain the optimized image. The experimental results show that the proposed algorithm is as good as other representative dehazing algorithms and enhance the middle and distant areas of the road image better. It effectively solves the color distortion and low contrast ratio of the near road surface and distant sky in the foggy road image, improves the visual effect of the reconstructed image, and has better image sharpening effect than dark channel prior algorithm, vision enhancement algorithm for homogeneous and heterogeneous fog, and typical dehazing algorithm based on deep learning.