On the estimation of face recognition system performance using image variability information

The type and amount of variation that exists among images in facial image datasets significantly affects Face Recognition System Performance (FRSP). This points towards the development of an appropriate image Variability Measure (VM), as applied to face-type image datasets. Given VM, modeling of the relationship that exists between the image variability characteristics of facial image datasets and expected FRSP values, can be performed. Thus, this paper presents a novel method to quantify the overall data variability that exists in a given face image dataset. The resulting Variability Measure (VM) is then used to model FR system performance versus VM (FRSP/VM). Note that VM takes into account both the inter- and intra-subject class correlation characteristics of an image dataset. Using eleven publically available datasets of face images and four well-known FR systems, computer simulation based experimental results showed that FRSP/VM based prediction errors are confined in the region of 0 to 10%