Single Stage Facial Recognition based on YOLOv5

Modern day facial recognition typically consists of four sequential stages: detection, alignment, representation and classification. Facial recognition and object detection are among the most widely researched fields in machine learning. However, they are often seen as entirely separate tasks. I propose a model that can handle facial recognition in just one stage, thus significantly improving the recognition speed. For this purpose, I have used the object detection model YOLOv5. YOLO, short for You Only Look Once, is known for being the fastest state-of-the-art object detection algorithm. With some key modifications to it’s design and loss function I have used it to train a model that can recognize faces from images containing one or more faces. For the experiment, I have used a custom dataset consisting of 283 images. The dataset has a mixture of individual and group photos, with some individuals repeating in some of the images. A total of 4 faces have been trained as individual classes in the modified YOLOv5 algorithm. When testing on new images, there is a significant increase in computation speed compared to other state-of-the-art facial recognition technologies that would be used for the same task, while having a marginal drop in accuracy. These are promising results that have potential applications in attendance systems and surveillance, for cases where the set of faces to be searched or recognized are known in advance.