Mutual features for robust identification and verification

Abstract

Noisy or distorted video/audio training sets represent constant challenges in automated identification and verification tasks. We propose the method of Mutual Interdependence Analysis (MIA) to extract "mutual features" from a high dimensional training set. Mutual features represent a class of objects through a unique direction in the span of the inputs that minimizes the scatter of the projected samples of the class. They capture invariant properties of the object class and can therefore be used for classification. The effectiveness of our approach is tested on real data from face and speaker recognition problems. We show that "mutual faces" extracted from the Yale database are illumination invariant, and obtain identification error rates of 2.2% in leave-one-out tests for differently illuminated images. Also, "mutual speaker signatures" for text independent speaker verification achieve state-of-theart equal error rates of 6.8% on the NTIMIT database.

Signature extraction using mutual interdependencies

“…GMIA(0) has already been tested on challenging real world applications such as illumination robust face recognition and textindependent speaker verification [5,6]. In this section, we evaluate the effect of l on the feature extraction and classification in both domains.…”

“…We tested the robustness to illumination scenarios of a GMIA(0)-based mutual face approach in Claussen et al [5]. In this problem, we have called the presumed common invariant feature ''mutual face''.…”

“…In this section, we evaluate the effect of l on the feature extraction and classification in both domains. First, we repeat the mutual face approach in Claussen et al [5] on the Yale face database [3] with GMIAðlÞ and compare the result for an optimized value of l to the previous GMIA(0) result. We aim to verify that both the results are similar for this illumination-invariant face recognition approach as indicated in Section 5.…”

“…Furthermore, Claussen et al [4][5][6] show how a mutual feature can be computed using mutual interdependencies in data (sounds and images) of the same class. Nonetheless, we aim for a deeper understanding of what GMIA features really represent, which lacks in previously published materials.…”

“…We aim to extract an invariant representation of high-dimensional instances of a single class, where the number of input instances N is smaller than their dimensionality D. An invariant is a property or feature of the input data that does not change within its class. Approaches that have been designed for this purpose are mutual interdependence analysis (MIA) and generalized MIA (GMIA) [4][5][6]. We revisit both methods in Sections 2 and 3, respectively, and parametrize GMIA with l, which subsumes MIA for l ¼ 0.…”