Face Hallucination Based on Independent Residual Features

A face hallucination scheme based on independent residual features (IRFs) is proposed in this paper. In the proposed scheme, a high-resolution (HR) face image is assumed as composition of two parts: an approximate and a residual image, where the approximate image is obtained by interpolating the corresponding low-resolution (LR) face image. According to this assumption, a residual image training database is established based on a set of training HR face images and their corresponding LR ones. And independent component analysis is used to extract IRFs from this established database. With constructing a cost function of least square, the IRFs are used to derive the corresponding residual image from any given LR face image and an expected HR one can be rendered. Experiments show the proposed scheme can effectively improve the resolution of a face image, on matter whether the face image belongs to the training face image set or not, under different light conditions.