[ Last changed: 17th October 1996 ]
Samples from stochastic signals with sufficient complexity need reveal only very little agreement in order to reject the hypothesis that they arise from independent sources. The failure of a statistical test of independence can thereby serve as a basis for recognising signal sources if they possess enough degrees of freedom. Combinatorial complexity of stochastic detail can lead to similarity metrics having binomial type distributions, and this allows decisions about the identity of signal sources to be made with astronomic confidence levels.
I will describe an application of these statistical pattern recognition principles in a system for biometric personal identification that analyses the random texture visible at some distance in the iris of a person's eye. There is little genetic penetrance in the phenotypic description of the iris, beyond colour, form and physiology. Since its detailed morphogenesis depends on the initial conditions in the embryonic mesoderm from which it develops, the iris texture itself is stochastic, if not chaotic. The recognition algorithm demodulates the iris texture with complex valued 2D Gabor wavelets, and coarsely quantises the resulting phasors to build a 256 byte `iris code' whose entropy is roughly 173 bits. Ergodicity and commensurability facilitate extremely rapid comparisons of entire iris codes using 32-bit XOR instructions. Recognition decisions are made by exhaustive database searches at the rate of about 10,000 persons per second.