Hallucination Of Face Images Across Multiple Modalities Using Tensors

Abstract

The existing face resolution techniques produces images with higher resolution from low resolution face images of a single modality. Single modality of facial images may be low resolution images of a fixed expression, fixed pose and a particular illumination. The technique presented here is hallucinating high resolution facial images over various modalities, i.e., with diverse facial expressions, poses and illumination conditions. This work addresses the following issues of the facial images taking various combinations of persons and expressions. 1. For a person with a single expression, is it possible to develop his other different expressions? 2. For a person with a single expression, is it possible to generate the same expression for different persons? 3. Given a facial image of low resolution, is it possible to recognize the person? To achieve the above issues, the work is concentrated on registering the unregistered raw images using an automatic face alignment algorithm and on developing a generalized tensor. The tensors are decomposed using HOSVD and finally synthesized to high or super resolution by Interpolation. This novel technique not only shows enhanced and superior performance over existing super resolution techniques but shows robustness in coping with hallucination of multi modal images under different practical conditions.

Keywords: HOSVD, Automatic face alignment algorithm, tensors