Edge-AI based Face Recognition System: Benchmarks and Analysis

The performance of real-time video processing tasks such as facial recognition has improved by many folds with the emergence of Artificial Intelligence (AI) and Convolutional Neural Networks (CNNs). However, face recognition models are computationally intensive and lack flexibility in deployment. Edge computing devices like Jetson (Nano, TX2) have overcome this gap by bringing high-speed and high-throughput computing capabilities to the edge. In this paper, we propose reliable integration of multi-face recognition system into compact and low-power edge devices. The deployed inference system uses a novel and light-weight face detector ‘FaceBoxes’, which is used for multi-face detection and extraction. ‘FaceNet’ is used as a face recognizer which outputs 128-dimensional embedding as feature vectors which is fed to a ‘Multilayer Perceptron (MLP)’ for face classification. We analyze the performance of facial recognition inference system on the two NVIDIA Jetson Edge computing boards (Nano, TX2). We also compare the performance of a TensorRT-based deep learning model optimization against a typical Tensorflow model implementation. The paper provides a detailed benchmarking analysis in terms of accuracy, FPS, execution time, memory usage and energy consumption. Additionally, the performance of the proposed face recognition system is compared with the state-of-the-art Multi-task Cascaded Convolutional Network (MTCNN) based face recognition system. Our findings can aid researchers in determining how this face recognition framework, when deployed on a particular Edge-GPU platform, might suit their needs for a specific face recognition application.