Face Recognition: A Comparison of Appearance-Based Approaches

We investigate the effect of image processing techniques when applied as a pre-processing step to three methods of face recognition: the direct correlation method, the eigenface method and fisherface method. Effectiveness is evaluated by comparing false acceptance rates, false rejection rates and equal error rates calculated from over 250,000 verification operations on a large test set of facial images, which present typical difficulties when attempting recognition, such as strong variations in lighting conditions and changes in facial expression. We identify some key advantages and determine the best image processing technique for each face recognition method. 1 Introduction Despite significant advances in face recognition technology, it has yet to be put to wide use in commerce or industry, primarily because the error rates are still too high for many of the applications in mind. These problems stem from the fact that existing systems are highly sensitive to environmental factors during image capture, such as variations in facial orientation, expression and lighting conditions. In this paper we attempt to address these issues by use of image pre-processing techniques, focusing on three face recognition methods, all coming under the general heading of appearance-based approaches: direct correlation; the eigenface method and the fisherface method. We begin with brief explanations of each face recognition method (section 2, 3 and 4), followed by a performance comparison of each system (section 5) with no image pre-processing (the ëbaseline systemsí). In Section 6 we outline a range of image pre-processing techniques, which may improve the baseline systems. By applying the face recognition methods to a substantial database of facial images (described in Section 7), producing graphs of FAR (False Acceptance Rate) against FRR (False Rejection Rate), from which the EER (equal error rate) is taken as a single comparative value, we compare the recognition accuracy across the full range of systems (Section 9).