Laplacian of smoothed image as representation for face recognition

This paper proposes the use of Laplacian of the smoothed image, called filtered image, for representation of an image for face recognition. The edge information is preserved around the zero-crossings of the filtered image. Smoothing of an image can be done either by the standard Gaussian filter or by the extension of the newly proposed concept of zero-frequency filter. The range of values of the filtered image are normalized for matching using a nonlinear transformation, without affecting the locations of the zero-crossings. The locality problem in matching of edge-based representation, is addressed by considering only the first few eigenvectors derived from the filtered images of the training set. The performance of the filtered image representation is shown to be significantly better than the representations based on the gray level or existing edge-based information.