A Linear Subspace Learning Approach via Low Rank Decomposition

Most existing subspace analysis methods for image recognition estimate the sample scatter matrices directly based on the training images. However, such methods do not consider the different contributions of different image components to image representation and recognition. Considering that the dominant component will contribute much more than the residual component to image pattern classification, we present a novel discriminate criterion which maximizes the total scatter of dominant components and minimizes simultaneously the total scatter of residual components. This criterion gives rise to a new subspace analysis method, namely the subspace analysis via low rank decomposition (SAL). Further, a supervised version of SAL (SSAL) is presented. The experimental results on benchmark image databases validated that SAL and SSAL outperform those representative subspace analysis methods such as PCA, LDA, LLP and SPP.