Discriminative Face Hallucination via Locality-Constrained and Category Embedding Representation

Recent years have witnessed the rapid development of face image hallucination techniques. However, previous methods are unsupervised and ignore label information training samples, leading to undesirable results. This article proposes a locality-constrained category embedding representation (LCER) method super-resolve in supervised manner by data representation. The proposed LCER incorporates locality prior into one unified framework, which aims learn both advantages preserving true typologic structure manifold discriminability exposing class subspace information. Such strategy allows not only preserve more sharpen details but also guarantee pattern be transferred mainly from same subject super-resolution reconstruction. Extensive experiments were conducted evaluate LCER, comparative results demonstrate that it achieved superior performance quantitative measurements visual impressions compared several state-of-the-art.