Lda with Subgroup Pca Method for Facial Image Retrieval

We need to capture properties of face image variations from training persons and generalize them to a new test person for robust image retrieval, which is essential for the case that there is only a single image of the test person to retrieval face images of the equal person from database. Conventional methods of Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) can be exploited to learn robust subspaces for face image variations contained in the training face images and the subspaces can be coped with the similar variations of new test persons. In this paper, a new method of LDA with subgroup PCA is proposed for an efficient learning of face image changes. Particularly the proposed method is designed to learn many kinds of extrinsic variations that are included in the training set. For this purpose the training set is partitioned into several subgroups by local property and PCA applied to the each subgroup yields the specialized subspace representation for the variations of the subgroups. On the contrary LDA is performed on the subgroup PCA features of the whole training set in order to avoid an overfitting under the influence of dominant variations. Several representations of the LDA with the subgroup PCA are finally merged by the weighted sum rule and it is shown to yield better recognition rate in the experimental results of the fluent face images collected from Purdue, PIE, ALTKOM, XM2VTS and BANCA databases.