Recognising persons by their iris patternsJohn Daugman
Professor of Computer Vision and Pattern Recognition
University of Cambridge, Computer Laboratory
Iris recognition is a biometric technology for identifying persons reliably by wavelet-encoding and analysis of the random patterns that are visible within the iris of an eye from some distance. Because the iris is a protected internal organ whose random texture is epigenetic and evidently stable, it serves as a kind of living password or key which has high entropy (about 250 bits in encodings of many databases). Recognition decisions are made with confidence levels high enough to support exhaustive searches very rapidly through national-sized databases. The high entropy is the origin of collision avoidance even in large databases. Today there are many public deployments of this technology around the world, mainly at border-crossings in lieu of passports, or in watch-lists, or entry control. Independent government tests (e.g. by NIST) confirm the extreme resistance to False Matches, and search speeds in the millions per second per CPU core.
The principle that underlies these recognition algorithms is the failure of an efficient test of statistical independence having many degrees of freedom, based on phase sequencing each iris pattern with quadrature 2D Gabor wavelets. Different eyes (including those of twins, or the right and left of one person) always pass this test of statistical independence, while images from the same iris almost always fail this test of independence, thereby signifying identity. Data used in this talk comes mainly from 200 billion iris cross-comparisons between different eyes, from a database consisting of 632,500 iris images acquired in the United Arab Emirates in a networked national border-crossing security programme that every day performs about 12 billion iris comparisons using these algorithms. Several other countries have now launched national biometric ID programmes, including UIDAI in India which is enrolling the iris patterns of all 1.3 billion citizens of India to secure benefits entitlements. The UIDAI performs hundreds of trillions of iris cross-comparisons every day, because of the need to detect duplicate identities as the database is built. Challenges for this technology include the difficulty of image capture, and the possibility of spoofing, which current research efforts seek to address.
A Cognitive Architecture for Object Recognition in VideoJosé Carlos Príncipe
Distinguished Professor of Electrical Engineering
Computational NeuroEngineering Laboratory (CNEL)
University of Florida
This talk describes our efforts to abstract from the animal visual system the computational principles to explain images in video. We develop a hierarchical, distributed architecture of dynamical systems that self-organizes to explain the input imagery using an empirical Bayes criterion with sparseness constraints and dual state estimation. The interpretation of the images are mediated through causes that flow top down and change the priors for the bottom up processing. We will present preliminary results and show simplified models for audition.
Clustering and Matching by Nonnegative Matrix FactorizationsErkki Oja
Professor of Computer Science and Engineering
Aalto University, School of Science and Technology
Many standard inference problems in computer vision involve combinatorial optimizations. A typical example is clustering in which some items, e.g. image pixels, are placed into groups in which the items within each group are more similar than items belonging to different groups. This is a central approach in image segmentation. Another example is graph isomorphism, or graph matching, where a permutation is sought for the node indices such that the two graphs become identical or closely similar. This is a central operation in scene understanding and object recognition. Mathematically, the solution to both problems is given by a binary orthogonal indicator matrix. The talk reviews a special relaxation approach where the binary matrix is replaced by a nonnegative approximately orthogonal continuous-valued matrix, and the problems can be presented in the form of Nonnegative Matrix Factorization (NMF), possibly with constraints.
Convergent versions of the learning rules are presented and experimental results are given in various inference problems.
ICIAR - International Conference on Image Analysis and Recognition
Send your comments to firstname.lastname@example.org