Face super-resolution via bilayer contextual representation

Abstract

Learning-based face super-resolution (SR), an important application field in computational imaging technology, offers a paradigm to enhance the quality and resolution of low-resolution (LR) facial images with the prior from training samples. The core of face SR is to learn the complex relationship between LR and high-resolution (HR) face images. Although most of the existing SR methods have achieved good results, the nonlinear nature of contextual prior is often ignored by locally linear assumption. In order to explore the accurate mapping relationship between LR and HR images, we propose a novel face SR method using bilayer contextual representation. In the first layer, we use context-patch to enrich the prior of image representation and shrink contextual dictionary in solution domain to induce local linear representation model in Euclidean space, thereby improving a set of LR images. In the second layer, kernel function provides a nonlinear mapping model for these enhanced images in Hilbert space, which maintains the consistency of the underlying manifold structure to capture a more accurate mapping between LR and HR. Furthermore, contextual residual learning is used to boost the reconstruction performance. The experimental results on FEI, CAS-PEAL-R1 and LDHF face databases show that the proposed approach preserves more image details and effectively improves the quality of the reconstructed image comparing with some state-of-the-art face SR methods.