Face hallucination via multiple feature learning with hierarchical structure

Abstract

In the past few years, neighbor-embedding (NE) based methods have been widely exploited for face hallucination. However, the existing NE based methods in spatial domain just employ single type of features for data representation, ignoring the compensatory information among multiple image features, resulting in bias in high resolution (HR) face image reconstruction. To tackle such problem, this paper presents a novel Multiple feature Learning model with Hierarchical Structure (MLHS) for face hallucination. Compared with conventional NE based methods, the proposed MLHS makes full use of multi-level information of face images, which can effectively remedy the flaw caused by just using single type of spatial pixel features, and adopts hierarchical structure to better maintain the manifold consistency hypothesis between the HR and low resolution (LR) patch spaces. The multiple learning strategy and hierarchical structure admit the proposed MLHS to well reconstruct the face details such as eyes, nostrils and mouth. The validity of the proposed MLHS method is confirmed by the comparison experiments in some public face databases.