Multi-biometric approaches to ear biometrics and soft biometrics

TLDR: This work develops methods that enable recognition using 2D ear images, and explores the growing field of ensemble biometrics, which subdivide a biometric feature into parts, and combine the results of several parts to yield recognition results.

Abstract: In this work, we explore hard and soft biometric systems. Hard biometrics are features that are used to uniquely identify individuals over time, while soft biometrics do not uniquely identify individuals and may not persist in the same state over an extended time. We develop methods that enable recognition using 2D ear images. This recognition is performed using a dataset which contains various lighting and pose conditions, as well as time lapse. We explore the growing field of ensemble biometrics, which subdivide a biometric feature into parts, and combine the results of several parts to yield recognition results. We vary the number of parts, the size of each part, and the method used to build each ensemble and report recognition improvements. We also allow the parts to change shape both before and during training, which further improves performance. We perform recognition using soft biometric features extracted from video. Although these features are not as reliable as traditional biometric features, they can still contribute to the recognition process. We find that using clothing color and height yield modest performance results that can be extended on their own or applied to other biometric systems.