Partial PCA in frequency domain

Partial principal component analysis has demonstrated to be a robust face recognition approach for big pose variations. The main idea behind P²CA is to use 3D (2.5D) data during the training stage but only a single 2D image in the recognition stage, to achieve a practical system for face recognition which is robust against pose variations. Nevertheles this approach present two major drawbacks: First one is that the training data should be well aligned to obtain good results; the second drawback is computational cost since the face space is computed maintaining the spatial dimensionality of the images. Thus, in this paper, we address both problems. First, we propose a two step procedure for a more accurate alignment of the 2.5D training texture maps. This alignment is done by applying a global and local affine transformation to all texture maps, while the local alignment is applied to some facial features using a triangulation mesh. Second, the computational cost is also reduced by applying P²CA in the frequency domain and perform the normalized cross correlation as similarity measure. Introducing these two improvements the recognition accuracy increases by 4% due to the alignment procedure and the computational time is reduced by a factor between 10 and 100 (depending on the number of eigenvectors used) due to the application of P²CA in the frequency domain.