Assessment of Time Dependency in Face Recognition

An important topic in face recognition is to determine the proper level of training—both in training set size and quality—for efficient learning of a face recognition algorithm. Another topic is the effect of time lapse between the target image and the query image on recognition performance. Both topics require a large set of face images where the same subjects have their images acquired repeatedly over a long period. This paper describes a long-term image acquisition project currently underway and discusses the results of face recognition experiments with respect to the training level and time delay on a large-scale data set. The experiments were performed by using the Principal Component Analysis approach. Results suggest that (a) increasing the size of the training data set improves the test performance, but the improvement levels off after some size, (b) performance degrades dramatically as the time lapse between the target image and the query image increases, and (c) performance figures reported in the literature based on images acquired on the same day may have little value in practice. We also address the potential problem of facial region extraction from the raw image. Index terms face recognition, biometrics, time-lapse performance, facial expression.