A multilinear collaborative representation preserving projections method for feature extraction

Abstract In this paper, a multilinear collaborative representation preserving projections (MCRPP) is developed for feature extraction and recognition on tensor objects. MCRPP is an unsupervised method built on collaborative representation and multilinear algebra. MCRPP calculates the projection matrix by simultaneously considering both the distribution information of all the samples and the spatial structure information of the sample. MCRPP aims to preserve the sparse reconstruction relations of data in the reduced subspace via collaborative representation reconstruction. Redefining the local multilinear scatter and the total multilinear scatter, MCRPP seeks the projection matrix by maximizing the ratio between the total multilinear scatter and local multilinear scatter using generalized eigenvalue decomposition. The final experimental results on FERET, AR and Yale face databases demonstrate that MCRPP is an efficient and effective method for feature extraction and dimensionality reduction.