融合暗通道先验损失的生成对抗网络用于单幅图像去雾

A generative adversarial network incorporating dark channel prior loss used for single image defogging

针对基于成对抗网络(GAN)的单幅图像去雾算法,其模型对样本真值过度拟合,而在自然图像上表现一般的问题,本文设计了一种融合暗通道先验损失的生成对抗网络来进行单幅图像去雾。该先验损失可以在网络训练中对模型预测结果产生影响,纠正暗通道特征图的稀疏性与偏度特性,提升去雾效果的同时阻止模型对样本真值过度拟合。另外,为了解决传统的暗通道特征图提取方法存在非凸函数,难以嵌入网络训练的问题,引入了一种基于像素值压缩的暗通道特征图提取策略。该策略将最小值滤波等效为对像素值压缩,其实现函数是一个凸函数,有利于嵌入网络训练,增强算法整体的鲁棒性。另外,基于像素值压缩的暗通道特征图提取策略不需要设置固定尺度提取暗通道特征图,对不同尺寸的图像均有良好的适应性。实验结果表明,相较于其它先进算法,本文算法在真实图像以及SOTS等合成测试集上均有良好的表现。

Single image defogging using generative adversarial networks(GAN)relies on annotated datasets,which is easy to cause over-fitting of ground truth,and usually performs not well on natural images.To solve this problem,this paper designed a GAN network incorporating dark channel prior loss to defogging single image.This prior loss can influence the model prediction results in network training and correct the sparsity and skewness of the dark channel feature map.At the same time,it can definitely improve the actual defogging effect and prevent the model from over-fitting problem.In addition,in order to solve the problem that the extraction method of traditional dark channel feature has non-convex function and is difficult to be embedded into network training,this paper introduces a new extraction strategy which compresses pixel values instead of minimum filtering.The implementation function of this strategy is a convex function,which is conducive to embedded network training and enhances the overall robustness of the algorithm.Moreover,this strategy does not need to set a fixed scale to extract the dark channel feature map,and has good adaptability to images with different resolutions.Experimental results show that the proposed algorithm performs better on real images and synthetic test-sets like SOTS when compared with other sota algorithms.