Skip to Main Content
Recently, two-dimensional locality preserving projections (2DLPP) has been receiving increasing attention for image analysis in both theory and applications. In this paper, we point out that the essential of 2DLPP is a special row-based locality preserving projections (LPP). So, 2DLPP can only extract features contained in row vectors of images, while the spatial arrangement information contained in column vectors, which is equally important for recognition problem, is completely discarded. To address this issue, we propose a new approach called structural two-dimensional locality preserving projections (S2DLPP) to fully extract features of both row and column vectors based on locality preserving criterion. S2DLPP is a manifold learning method that identifies local structure information rather than only row information as in 2DLPP, which makes S2DLPP more accurate in finding discriminative information. Like 2DLPP, S2DLPP is formulated as solving a generalized eigenvalue problem, which is computationally straightforward and does not involves singularity. Experiments on handwritten digit recognition and face recognition demonstrate the effectiveness of the proposed method.