Abstract
A fast and accurate stand-alone face recognition system is crucial for surveillance purposes; however, it is also important to keep the costs as low as possible. Herein, we address this issue by proposing a preliminary analysis of embedded Machine Learning techniques by using Erdős-Rényi sparse random networks. The idea is to develop a compact and reliable ANN to conduct a multi-class classification of low-resolution face images to simulate the scenario of having cheap security cameras with an embedded ANN. The study considered two architectures (ResNet, and AlexNet inspired CNNs) with a sparsity level varied up to 90%. To achieve comparable results, the image resolution was varied from 32×32 up to 96×96. The analyses unveiled that for low-resolution images, the best trade-off between accuracy and sparsity level has been achieved with ResNet architectures and a sparsity level of 70% outperforming the benchmark (i.e., with no sparsity).