Abstract Representation learning methods have attracted considerable attention for learning-based face super-resolution in recent years. Conventional methods perform local models learning on low-resolution (LR) manifold and face reconstruction on high-resolution (HR) manifold respectively, leading to unsatisfactory reconstruction performance when the acquired LR face images are severely degraded (e.g., noisy, blurred). To tackle this issue, this paper proposes an efficient multilayer locality-constrained matrix regression (MLCMR) framework to learn the representation of the input LR patch and meanwhile preserve the manifold of the original HR space. Particularly, MLCMR uses nuclear norm regularization to capture the structural characteristic of the representation residual and applies an adaptive neighborhood selection scheme to find the HR patches that are compatible with its neighbors. Also, MLCMR iteratively applies the manifold structure of the desired HR space to induce the representation weights learning in the LR space, aims at reducing the inconsistency gap between different manifolds. Experimental results on widely used FEI database and real-world faces have demonstrated that compared with several state-of-the-art face super-resolution approaches, our proposed approach has the capability of obtaining better results both in objective metrics and visual quality.

中文翻译：

---

构建噪声鲁棒人脸超分辨率的多层局部约束矩阵回归框架

摘要 近年来，基于学习的人脸超分辨率表示学习方法引起了广泛关注。传统方法分别在低分辨率（LR）流形上执行局部模型学习和在高分辨率（HR）流形上进行人脸重建，当获取的 LR 人脸图像严重退化（例如，嘈杂、模糊）时，导致重建性能不理想。为了解决这个问题，本文提出了一种高效的多层局部约束矩阵回归（MLCMR）框架来学习输入 LR 补丁的表示，同时保留原始 HR 空间的流形。特别，MLCMR 使用核范数正则化来捕捉表示残差的结构特征，并应用自适应邻域选择方案来找到与其邻居兼容的 HR 块。此外，MLCMR 迭代地应用所需 HR 空间的流形结构来诱导 LR 空间中的表示权重学习，旨在减少不同流形之间的不一致差距。在广泛使用的 FEI 数据库和真实世界人脸的实验结果表明，与几种最先进的人脸超分辨率方法相比，我们提出的方法能够在客观指标和视觉质量方面获得更好的结果。MLCMR 迭代地应用所需 HR 空间的流形结构来诱导 LR 空间中的表示权重学习，旨在减少不同流形之间的不一致差距。在广泛使用的 FEI 数据库和真实世界人脸的实验结果表明，与几种最先进的人脸超分辨率方法相比，我们提出的方法能够在客观指标和视觉质量方面获得更好的结果。MLCMR 迭代地应用所需 HR 空间的流形结构来诱导 LR 空间中的表示权重学习，旨在减少不同流形之间的不一致差距。在广泛使用的 FEI 数据库和真实世界人脸的实验结果表明，与几种最先进的人脸超分辨率方法相比，我们提出的方法能够在客观指标和视觉质量方面获得更好的结果。