Abstract
This thesis presents an improved two-step supervised face hallucination framework termed from the input low-resolution (LR) face image to the high-resolution (HR) image. To solve the special facial problem, we propose a novel face hallucination using Bayesian global estimation, local basis selection with support vector machine (SVM) and label information to achieve supervised learning for constructing super-resolution (SR) frontal images from the input LR face image. This proposed framework mainly consists of two steps: the global estimation step and the local facial-parts refinement using selection local bases selection step. In order to improve the face hallucination performance, we further employ face recognition (SVM) to find the similar face structure bases (global and local bases) as an input face image. In the global estimation, we use face recognition to select global/PCA bases and adopt a maximum a posteriori (MAP) estimator to estimate the optimum set of coefficients in the low-dimensional domain for hallucinating HR face image via a linear combination of the global bases. In the local refinement step, we use face recognition to select local/overcomplete nonnegative matrix factorization (ONMF) bases to refine the facial parts (i.e. eyes, nose and mouth). Experimental results show that our improved framework can effectively enhance visual effects and demonstrate that the good performance of our approach with face recognition is justified in that our reconstruction results are better than those produced by the other hallucination methods, such as visual quality and objective quality assessment (face recognition).