Unidirectional variation and deep CNN denoiser priors for simultaneously destriping and denoising optical remote sensing images
Unidirectional variation and deep CNN denoiser priors for simultaneously destriping and denoising optical remote sensing images
复制标题
用于同时对光学遥感图像进行去条纹和去噪的单向变化和深度 CNN 去噪器先验
DOI：
10.1080/01431161.2019.1580821
复制
发表时间：
2019-08-03
影响因子：
3.4
通讯作者：
Xiong, Shiqi
中科院分区：
文献类型：
--
作者：
Huang, Zhenghua;Zhang, Yaozong;Xiong, Shiqi
条纹和随机噪声是光学遥感图像中常见的两种不同的退化现象，它们通常分别被建模为逆问题。在求解这些逆问题时，基于模型的优化方法和判别学习方法是目前比较流行的两种方法，但它们各有优缺点，例如，基于模型的优化方法灵活，但通常耗时长，而判别学习方法测试速度快，但受任务的特殊性限制。为了提高测试速度并获得良好的性能，本文将深度卷积神经网络（DCNN）去噪先验器集成到单向变化（UV）模型中，命名为UV-DCNN，对光学遥感图像进行去条纹和去噪同时进行。所提出的UV-DCNN方法可以通过交替最小化优化方法有效地求解。定量和定性实验结果都证明了该方法的有效性，甚至优于目前的方法，其令人满意的计算时间使其适合广泛应用。
ABSTRACT Stripe and random noise are two different degradation phenomena commonly co-existing in optical remote sensing images, which are often modelled as inverse problems, respectively. When solving those inverse problems, model-based optimization and discriminative learning methods are fashionably employed but have their respective merits and drawbacks, e.g., model-based optimization methods are flexible but usually time-consuming while discriminative learning methods have fast testing speed but are limited by the specialized task. To improve testing speed and obtain good performance, this paper integrates deep convolutional neural network (DCNN) denoiser prior into unidirectional variation (UV) model, named as UV-DCNN, to simultaneously destripe and denoise optical remote sensing images. The proposed UV-DCNN method can be efficiently solved by the alternating minimization optimization method. Both quantitative and qualitative experiment results validate that the proposed method is effective and even better than the state-of-the-arts, its satisfactory computation time makes it suitable for extensive application.