GPU-Based Multiple Face Recognition Using YOLOv5x

The growth of deep learning has led to impressive achievements and challenging issues in artificial intelligence and computer vision fields. Face recognition based on convolutional neural networks (CNNs) is one of the challenging and interesting hot topics in these fields. Recently, YOLO (You Only Look Once) became one of the more robust and attractive frameworks adopted for object detection in real-time. Despite that it is originally designed for object detection, the work of this paper is succeeded to apply it for face recognition with multiple faces recognized simultaneously in real time. The last version of this platform, known as YOLOv5x, is adopted to satisfy the desired task based on embedded GPU (Graphics Processing Unit) edge device. The proposed system is trained using custom-created dataset of 7378 augmented images, tested, and validated using Google Colab Notebook and Jetson Nano developer board. This board is supported by a parallel computing platform called CUDA (Compute Unified Device Architecture) enhanced with PyTorch framework for real-time acceleration without using DeepStream real-time server or Docker container. The developed system can recognize up to 20 faces simultaneously with high inference time of 0.817 s and precision of 99.8%.