Research Article
An Improved EMD-Based Dissimilarity Metric for Unsupervised Linear Subspace Learning
Algorithm 1
The improved EMD metric for LPP.
Input: the sample set with samples, parameter , block parameter , pHOG bins, nearest neighbors parameter | Output: adjacency graph , weight matrix , transformation matrix , eigenvalues , and subspace y | Whileāā | Extract HOG histogram over each block of per face image | Carry out the pooling operation over each block and then get the pHOG histogram | Obtain the grids of pHOG vector for one face image and grids of pHOG vectors for the rest of face images , | Compute the dissimilarity between and by Equations (8) and (9) | Obtain the nearest neighbors of the face image : | EndWhile | Build the adjacency graph and calculate the corresponding weight matrix by Equation (10) | Begin // compute the projection | Get the diagonal matrix | Solve the generalized eigenvector problem of Equation (11) on the sample set | Get the eigenvectors with respect to eigenvalues | End // compute the projection | Obtain the transformation matrix | Obtain the subspace for the sample set by Equation (13) | (16) Perform face recognition by the classifier |
|