Manifold learning-based sample selection method for facial image super-resolution

Learning efficiency, related to not only the size of training set but also to the usage of samples, is an important issue of learning-based super-resolution (SR). We propose a face hallucination method with adaptive sample selection ability by learning from the facial manifold. Employing locality preserving projections (LPP) to analyze the intrinsic features on the local facial manifold, our method searches out image patches dynamically in the LPP subspace, which makes the training set tailored to the input patch. Using the selected training set, we develop a patch-based eigen-transformation algorithm to efficiently restore the lost high-frequency components of the low-resolution face image. Experiments on synthetic and real-life images fully demonstrate that the proposed adaptive sample selection SR method can achieve better performance than some state-of-the-art learning-based SR techniques with less computational cost by utilizing a relative small sample, especially under the case of low quality input images. (C) 2012 Society of Photo-Optical Instrumentation Engineers (SPIE). [DOI: 10.1117/1.OE.51.4.047003]