Locally Salient Feature Extraction Using ICA for Content-Based Face Image Retrieval

The paper focuses on face image retrieval based on higher level statistical features. The principal independent content feature (PICF) is extracted by independent component analysis (ICA) to represent facial images, and a corresponding similarity measurement is employed. The PICF method encodes face images with locally salient features from a set of training images, which operates in a reduced principal component analysis (PCA) space, and an enhanced retrieval is achieved by using the similarity measurement after the two-stage-retrieving. The simulation is performed by using the ORL database, where the face images vary in illumination, expression, pose, and scale. The results show that the PICF method using the corresponding similarity measurement has better retrieving performance than classical PCA or ICA method with the usual measurement, and the two-stage-accuracy can reach 97.5%