Face retrieval by an adaptive Mahalanobis distance using a confidence factor

This paper proposes an adaptive Mahalanobis distance for face retrieval. The distance is derived from a posterior distribution of observation errors in features categorized by con dence of face images. Since the distance is calculated considering error variances of each dimension according to the con dence, it can re ect error distribution of each matching more precisely than a standard Mahalanobis distance. We apply this distance to eigenface techniques using image contrast and asymmetric components of face images as the con dence. To evaluate our proposed distance in face retrieval, we made experiments using MPEG-7 face descriptors as eigenface features. The best match ratio was improved from 93.5% to 97.6% compared with the weighted distance described in MPEG-7 by using the proposed distance.