| Popis: |
The recent growth in CCTV systems and the challenges of automatically identifying humans under the adverse visual conditions of surveillance have increased the interest in soft biometrics, which are physical and behavioural attributes that are used to semantically describe people. Soft biometrics enable human identification under the challenging conditions of surveillance where it is impossible to acquire traditional biometrics such as iris and fingerprint. The existing work on facial soft biometrics is focused on categorical attributes, while comparative attributes have received very little attention, although they have demonstrated a better accuracy. Thus, it is still unknown whether comparative soft biometrics can scale to large and more realistic databases. Also, the automatic retrieval of comparative facial soft biometrics from images needs to be investigated. The purpose of this thesis is to explore human identification and verification in large and realistic databases via comparative facial soft biometrics using the Labelled Faces in the Wild (LFW) database. A novel set of comparative facial soft biometrics is introduced, and a thorough analysis that assesses attribute significance and discriminative power is presented. Also, a set of identification and verification experiments was conducted to evaluate the comparative facial soft biometrics. Moreover, this thesis proposes MIURank, a novel fully unsupervised ranking algorithm that is based on mutual information. The experiments demonstrate that a correct match can be found in the top 71 retrieved subjects from a database of 4038 subjects by comparing an unknown subject to ten subjects only. Additionally, the experiments reveal that face retrieval by verbal descriptions in a database of images can yield a correct match in the top 15 retrieved subjects from a database of 430 subjects. Furthermore, the performance analysis of the MIURank algorithm shows that it can result in a ranking accuracy that is comparable to the maximum likelihood estimator of Bradley-Terry and the state-of-the-art SerialRank algorithm. By these analyses and developments, it is now possible not only to use human labels for recognition, but also to derive them by computer vision. |
