Learning binary codes for effective and efficient search of multibiometric features

Large-scale personal identification systems use multiple biometric traits (e.g. face, fingerprint, iris) from the same person, so-called multibiometrics, to increase the recognition accuracy and population coverage. Core to such systems is the search of multibiometric identities in large-scale databases. With ever-increasing data volume and access demand, it is necessary to develop efficient search methods subject to a certain accuracy level. This is particularly important for biometric identity management in critical national security applications that involve large scale computationally intensive tasks, such as identity de-duplication and secured identification.

An efficient search method consists in narrowing down the search range and reducing the matching complexity. The basic idea of biometric indexing is to assign an index vector to every identity in the database, with the aim of retrieving a small number of candidate identities for matching. Earlier work in this area has been focused on single modality indexing schemes, which typically derive the index values by dimension reduction of the respective biometric features. Multibiometric indexing schemes are emerging, but existing techniques rely on modality-specific matchers and matching references that are data-dependent.

Moreover, indexing methods based on real-valued index descriptors of high-dimension biometric features often require significant computational efforts and memory storage. Binary representations can ensure fast operations and low storage costs. However, biometric features are in general not represented in binary form. Known binarization schemes for biometric features were designed for one-to-one verification (e.g. in biometric cryptosystems) that requires accurate matching. The resulting binary templates typically have long bit-length, which is problematic for large scale searches.

This project seeks to develop binary index codes for effective and efficient search of multibiometric features. Our approach is to leverage machine learning techniques to integrate binary embedding with nearest neighbour search procedures. We design learning objectives to obtain compact binary representations by exploiting the feature distribution in the original biometric feature space. The binary index codes learned from heterogeneous biometric features will be combined systematically into a compact form for indexing multibiometric identities in large-scale databases.