Key words
dynamic biometric images, estimation of informative features, artificial neural networks, statistical functionals, correlation between features.
articles
The author:
Sulavko A. Published in
№ 1(73)
27 february 2018 year
Rubric: Models and Methods
Rubric: Models and Methods
Testing of neurons based on statistical functionals for verifying biometric images in feature space with different informativeness
The scale of features informativeness is formulated and presented. The efficiency of «wide» neural
network neurons based on the various variations of following criteria for checking the distribution law
of a random variable was evaluated: Smirnov-Kramer-von Mises, Anderson-Darling, Watson, Frotsini,
average geometric comparison functions for probability densities Kolmogorov-Smirnov, Cooper. A criterion
for the maximum area of intersection of the comparable probability density functions is proposed.
Variants for the modernization of functionals for the processing of features with a noticeable and high
correlation dependence are found, in particular, on the basis of the Smirnov-Kramer-von Mises criterion.
The analysis of the results of the work made it possible to determine the most suitable conditions
for the use of the functionals under consideration, depending on the features of the feature space. The
boundary conditions of the experiment correspond to the features of dynamic biometric patterns. First
of all, the research was aimed at solving the problems of recognizing secret biometric images, but the
results obtained can be used in any verification task, where there are dependent and independent signs
of average informativeness, as well as little informative and extremely little informative features with
a distribution of values close to normal.
The author:
Sulavko A.
Degree:
Cand. Sci. (Eng.), Assistant Professor, Integrated Information Security Department, Omsk State Technical University
Location:
Omsk, Russia
