A Robust Dimensionality Reduction Method from Laplacian Orientations

Most dimensionality reduction methods are usually based on dissimilarity measurement of pixel intensities which can not obtain a more robust dissimilarity measurement. To address this problem, in this paper, we propose a novel robust dimensionality reduction method from Laplacian orientations. This method does not directly manipulate pixel intensity, which introduces Laplacian orientations, combined with the kernel method, and ultimately robust dimensionality reduction. The use of the Laplacian orientations results in a more robust dissimilarity measurement between images. Our method is as simple as standard intensity-based learning, yet much more powerful for efficient dimensionality reduction method. Our experiments show that the proposed method for different expressions, different illumination conditions and different occlusions under different illumination conditions has better robustness, and achieves a higher recognition rate. For a single sample per person, the proposed algorithm can also obtain a higher recognition rate.