Optimization of Convolutional Neural Networks for Face Recognition Problems Aimed to Resources Constrained Platforms

In this paper is proposed a method to obtain high recognition rates of face recognition problems using low complexity convolutional neural networks, tailored for resource-constrained computing platforms. Different variants of CNN architectures were considered and thoroughly evaluated on datasets such as YouTube Faces Database and Aligned Face Dataset for Face Recognition. We experiment by training our architectures on various computing platforms including CPU, GPU and evaluate performances such as recognition rates, prediction estimation, throughput and latency. In optimizing the specific parameters, a trade-off between a good accuracy and a low complexity was considered.