一种基于改进AOD-Net的航拍图像去雾算法

针对航拍图像易受雾气影响, AOD-Net (All in one dehazing network)算法对图像去雾后容易出现细节模糊、对比度过高和图像偏暗等问题, 本文提出了一种基于改进AOD-Net的航拍图像去雾算法. 本文主要从网络结构、损失函数、训练方式三个方面对AOD-Net进行改良. 首先在AOD-Net的第二个特征融合层上添加了第一层的特征图, 用全逐点卷积替换了传统卷积方式, 并用多尺度结构提升了网络对细节的处理能力. 然后用包含有图像重构损失函数、SSIM (Structural similarity)损失函数以及TV (Total variation)损失函数的复合损失函数优化去雾图的对比度、亮度以及色彩饱和度. 最后采用分段式的训练方式进一步提升了去雾图的质量. 实验结果表明, 经该算法去雾后的图像拥有令人满意的去雾结果, 图像的饱和度和对比度相较于AOD-Net更自然. 与其他对比算法相比, 该算法在合成图像实验、真实航拍图像实验以及算法耗时测试的综合表现上更好, 更适用于航拍图像实时去雾.

A Defogging Algorithm for Aerial Image With Improved AOD-Net

As the aerial images are easily affected by fog, the defogging image processed by AOD-Net is prone to problems such as blurring of image details, excessive contrast and low brightness. A defogging algorithm for aerial image with improved AOD-Net was proposed. We mainly improve AOD-Net from three aspects: network structure, loss function and training method. Firstly, we add the feature image of the first layer to the second feature fusion layer of AOD-Net, the traditional convolution method is replaced by the fully point-wise convolution, and the multi-scale structure is used to enhance the ability of the network to deal with details. Then, in this paper, a composite loss function including image reconstruction loss function, SSIM loss function and TV loss function is used to optimize the contrast, brightness and color saturation of the defogging image. Finally, we use a segmented training method to further improve the quality of the defogging image. The experimental results show that the image defogged by the proposed algorithm has satisfactory defogging results, the saturation and contrast are more natural than AOD-Net. Compared with other comparison algorithms, the proposed algorithm has better comprehensive performance in synthetic image experiments, real aerial image experiments and time-consuming tests, and is more suitable for real-time defogging of aerial images.