OpenCL implementations of a genetic algorithm for feature selection in periocular biometric recognition

Abstract

This paper explores OpenCL implementations of a genetic algorithm used to optimize the features vector in periocular biometric recognition. Using a multi core platform the algorithm is tested for CPU and GPU, exploring different parallelization levels for each operator of the genetic algorithm. The results show that using the GPU platform it is possible to accelerate the algorithm by several orders of magnitude, with a recognition rate similar to the one obtained in the sequential version. The results also show that it is possible to use only a small portion of the features without any degradation of the classifier's recognition rate. © 2012 Springer-Verlag.