Pawel Drozdowski defended his Master's Thesis on 'Efficient Biometric Identification in large-scale Iris Databases'

The growing interest and acceptance of biometrics has resulted in an increasing number and size of biometric systems around the world. These systems use measurable and distinctive human characteristics for, among others, the purpose of automatised recognition of individuals. Several country-wide deployments, such as the Indian AADHAAR project, were spawned in recent years. The daily operation of these systems faces an immense computational load, which can contribute to increased system costs and reduced system usability. The goal of this thesis is to perform research in the area of biometric workload reduction in identification scenarios for large (human) iris databases. The iris has been selected as the biometric characteristic for the project, due to it being well-suited for use in large systems and is commonly used in actual deployments around the world.

The research in this thesis was carried out using a recently proposed, novel biometric indexing approach based on Bloom filters and binary search trees. During the course of this thesis, said approach was thoroughly analysed and expanded upon. In particular, several key improvements that ensure the system’s scalability were proposed, implemented and tested quantitatively in terms of biometric performance and workload reduction.
The system was shown to achieve an excellent biometric performance and a substantial workload reduction on a dataset of images with low intra-class variation. It was also discovered, that the biometric performance was severely impaired in the tests on a dataset of images with high intra-class variation. The results suggest that the approach is fully scalable in terms of enrolled database size. The biometric sample quality, however, may be a limiting factor. Furthermore, a brief investigation into multi-iris indexing has been conducted and shows great promise for future research. It is, best to this author’s knowledge, the first such study in the scientific literature.

In addition to the empirical testing, a general statistical model for Bloom filter based biometric indexing was presented. Based on several variables, the model allows for a theoretical assessment of the viability of a system configuration. Lastly, due to the incomprehensibility of current biometric workload reduction result reporting methods in the scientific literature, a unified and transparent methodology of result reporting was proposed. The aim of this undertaking was to elucidate this important issue and to serve as a basis for a better scientific process henceforth.