Face virtual pose generation using multi resolution subspaces

In this paper a new method for face virtual pose generation is presented. The proposed method uses subspace image representation. The general problem of blurring is addressed by introducing a multi resolution time-frequency analysis to subspace image representation. The training gallery contains face images in two different poses. Undecimated Wavelet Transform is applied on all training face images in first pose and the corresponding images in the second pose to produce image subbands. Then, a new subspace is constructed for each subband in both poses. The mapping between two corresponding subbands of two poses is learnt using linear regression. The resulted mapping matrix is used to generate virtual pose of a new given face. Experimental results with ICA as the subspace method on CMU PIE database show that the proposed method has a better performance on complex areas of face such as eyes, eyebrows and noses compared to global regression methods such as PCA and state-of-the-art methods such as Locally linear regression (LLR).