Réalisation de classificateurs neuroniques pour la reconnaissance à basse puissance de visages dans des séquences vidéo

Abstract

Biometrics have been suggested as a solution to the multiplication of codes and passwords necessary to secure mobile electronic devices. In this document, a technique for the realisation of a low-power video-based facial recognition system is proposed and the performance of three hardware implementations is evaluated. For an efficient implementation, the training of fuzzy ARTMAP neural classifiers with a training strategy to simultaneously optimize the hyper-parameters, the synaptic weight values and the neural configuration of the networks through multi-objective particle swarm optimization. Three commercially available processors have been selected to measure the performance of hardware implementation of the system. They are the Core i3-530 from Intel, the Atom N270 also from Intel and the Vocallo MGW from Octasic. The quality of the identity predictions, the memory usage, the processing time and the energy consumption has been evaluated using the IIT-CNRC video database. Results have shown that the time spent communicating between processing cores is negligible compared to the time spent calculating the identity predictions when using the Vocallo MGW. It is possible to estimate the classification rate of a fuzzy ARTMAP neural classifier trained with the MOPSO training technique for a given numbers of characteristics kept in the input patterns by using a nearest neighbor type classifier. The exponential growth of the network’s memory usage when more characteristics are kept suggests using the least amount of characteristics possible. Finally, the use of the fuzzy ARTMAP neural classifier trained with the MOPSO strategy allows finding the best trade-off between the quality of the predictions made and the amount of resources used. This solution paired to a multi-core low power processor provides low energy consumption while maintaining a good classification rate and a processing time close to real-time.

Date	11 Jun 2012
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Éric Granger (Supervisor) & Claude Thibeault (Co-supervisor)