Abstract:
Convolutional Neural Networks have proved an excellent efficiency in several modern applications, for example, systems of face identification, like VGGFace and DeepFace, have achieved an excellent performance. However, these models still require huge memory for computation that involves an expensive computation cost especially for applications that use huge databases and that are running in embedded devices. To deal with this problem, we propose in this paper a hybrid identification system that is based on the compression of the VGGFace model for the feature extraction step, and on the indexation and the parallelization for the identification task. The proposed system has been evaluated in term of the Rank-1 prediction, the time of identification and the speed-up using two public face databases. Our experimental results illustrate the ability of the proposed system to preserve the performance while keeping a reasonable time of identification compared to two face identification systems based, respectively, on the original VGGFace model and on the Inception V3 model.