Robust face super-resolution via position-patch neighborhood preserving

Abstract

By incorporating the priors that human face is a class of highly structured object, position-patch based face hallucination methods represent the test image patch through the same position patches of training faces by employing least square estimation or sparse coding. Due to they cannot provide unbiased approximations or ignore the influence of spatial distances between the test image patch and training basis image patches, the obtained representation is not satisfactory. In this paper, we propose a simpler yet more effective scheme called Position-patch Neighborhood Preserving (PNP). We improve existing SR methods by exploiting locality constraint and shrinkage measures to maintain locality and stability simultaneously. Moreover, our method use less similar patches, face hallucination is fast and robust. Various experimental results on standard face database show that our proposed method outperforms state-of-the-art methods in terms of both objective metrics and visual quality.