Tensor-based subspace learning and its applications in multi-pose face synthesis

Abstract

Facial pose synthesis is applied to generate much required information for several applications, such as public security, facial cosmetology, etc. How to synthesize facial pose images from one image accurately without spatial information is still a challenging problem. In this paper we propose a tensor-based subspace learning method (TSL) for synthesizing human multi-pose facial images from a single two-dimensional image. In the proposed TSL method, two-dimensional multi-pose images in the database are previously organized into a tensor form and a tensor decomposition technique is applied to build projection subspaces. In synthesis processing, the input two-dimensional image is first projected into its corresponding projection subspace to get an identity vector and then the identity vector is used to generate other novel pose images. Our technique is applied on KAO-Ritsumeikan Multi-angle View, Illumination and Cosmetic Facial Database (MaVIC) and experiment results show the effectiveness of our proposed method for facial pose synthesis.