Abstract Because of the ill-posedness of multi-frame super resolution (MSR), the regularization method plays an important role in the MSR field. Various regularization terms have been proposed to constrain the image to be estimated. However, artifacts also exist in the estimated image due to the artificial tendency in the manually designed prior model. To solve this problem, we propose a novel regularization-based MSR method with learned prior knowledge. By using the variable splitting technique, the fidelity term and regularization term are separated. The fidelity term is associated with an “ L 2 {L^{2}} - L 2 {L^{2}} ” form sub-problem. Meanwhile, the sub-problem respect to regularization term is a denoising problem, which can be solved by denoisers learned from a deep convolutional neural network. Different from the traditional regularization methods which employ hand-crafted image priors, in this paper the image prior model is replaced by learned prior implicitly. The two sub-problems are solved alternately and iteratively. The proposed method cannot only handle complex degradation model, but also use the learned prior knowledge to guide the reconstruction process to avoid the artifacts. Both the quantitative and qualitative results demonstrate that the proposed method gains better quality than the state-of-the-art methods.

中文翻译：

---

通过深度即插即用 CNN 正则化实现多帧超分辨率

摘要 由于多帧超分辨率（MSR）的不适定性，正则化方法在MSR领域发挥着重要作用。已经提出了各种正则化项来约束要估计的图像。然而，由于人工设计的先验模型中的人为趋势，估计图像中也存在伪影。为了解决这个问题，我们提出了一种新的基于正则化的 MSR 方法，该方法具有学习的先验知识。通过使用变量分裂技术，保真项和正则化项被分离。保真度项与“L 2 {L^{2}} - L 2 {L^{2}}”形式的子问题相关联。同时，关于正则化项的子问题是一个去噪问题，可以通过从深度卷积神经网络中学习的去噪器来解决。与传统的使用手工图像先验的正则化方法不同，本文将图像先验模型隐式地替换为学习先验。这两个子问题交替和迭代地解决。所提出的方法不仅可以处理复杂的退化模型，而且还可以使用学习到的先验知识来指导重建过程以避免伪影。定量和定性结果都表明，所提出的方法比最先进的方法获得了更好的质量。但也使用学习到的先验知识来指导重建过程以避免伪影。定量和定性结果都表明，所提出的方法比最先进的方法获得了更好的质量。但也使用学习到的先验知识来指导重建过程以避免伪影。定量和定性结果都表明，所提出的方法比最先进的方法获得了更好的质量。