Abstract:

The low frequency subimage is  usually used for face recognition based on wavelet transform (WT) methods. However, some important information hidden in other high frequency subimages will be  unavoidably lost. To solve this problem, two methods were  presented for face recognition by ensemble of wavelet subspaces, and the comparisons with other related methods were put forth by experiments. In the first method, the wavelet low frequency subimages at each layer were integrated for face recognition. In the second method, face images were first  decomposed into different subimages with L layer wavelet transform, and then L wavelet subspace images were obtained by averaging three high frequency subimages of each layer and integrating the low frequency subimage of each layer. Finally the L wavelet subspace images were integrated for face recognition. The proposed methods could make full use of the information provided by the different frequency subimages and the accuracy of face recognition was improved. The experimental results of three face databases (ORL, YALE, and JAFFE) showed that the proposed methods, especially the second method, could obtain a higher accuracy than other related methods.

Key words:  face recognition, wavelet transform, ensemble subspaces, 2D principal component analysis(2DPCA), 2D linnear discriminant analysis(2DLDA)