Hallucinated 3d face model from a single 2d low-resolution face using machine learning

Abstract

In this paper, the 3D face hallucination system is proposed on both 2D training face images as well as respective 3D training face models with grey-level. The proposed method hallucinates the 3D high-resolution model patch using same position of each image patches of interpolated 2D training image and 3D high-resolution training face model for low-resolution input image. Firstly, the optimal weights of the 2D input low-resolution image position-patches are estimated with the corresponding 2D low-resolution training image patches. The canonical correlation analysis (CCA) is used to learn the mapping between the 2D interpolated face training image and the 3D face model with respect to their weights. Secondly, the corresponding 3D face model patch with weight by matching high score among the 2D interpolated training image patches and 3D training face model is selected. Finally, the 3D high-resolution facial model is formed by integrating the hallucinated 3D patches which are obtained through mapping patches with respective weights. In order to evaluate the performances of the above approaches, we used example based learning methods to obtain the high-resolution output for a low-resolution input. In this approach, we used the available frontal data sets such as FERET, CAS-PERL and CMU to analysis the performance and some parameters are also considered, which may affect the results from the above proposed method.