Maximum margin criterion (MMC) is a well-known method for featureextraction and dimensionality reduction. However, MMC is based onvector data and fails to exploit local characteristics of imagedata. In this paper, we propose a two-dimensional generalizedframework based on a block-wise approach for MMC, to deal withmatrix representation data, that is, images. The proposed method,namely, block-wise two-dimensional maximum margin criterion(B2D-MMC), aims to find local subspace projections usingunilateral matrix multiplication in each block set, such that inthe subspace a block is close to those belonging to the same classbut far from those belonging to different classes. B2D-MMC avoidsiterations and alternations as in current bilateral projectionbased two-dimensional feature extraction techniques by seeking aclosed form solution of one-side projection matrix for each blockset. Theoretical analysis and experiments on benchmark facedatabases illustrate that the proposed method is effective andefficient.
展开▼
