关键词: 人脸识别, RetinaFace, FaceNet, RKNN模型转换

Abstract: During the epidemic outbreak, wearing masks has become an important public epidemic prevention measure. However, masks will lead to a reduction in recognition accuracy and a slowdown in detection speed of image-based face recognition systems. This study designs a masks-wearing face recognition system based on RetinaFace and FaceNet algorithms. First, the RetinaFace algorithm
is used to train the RetinaFace model to realize the face key point positioning function. Secondly, the FaceNet algorithm is used to train
the FaceNet model to extract the face feature vector and build a feature database. Finally, by comparing a Euclidean distance between
the facial feature vector to be recognized and the vectors in the database, the recognition result is output. The experimental results show that the recognition speed of the system after RKNN model conversion exceeds 25 f/s. On the dataset Mask-LFW, when the Euclidean distance threshold between feature vectors is set to 1.01, the ACC is up to 93.78%, and the AUC is 91.03%. The system meets realtime requirements and has high detection accuracy, which can be used in laboratories, company offices, factories and some public places.

Key words: face recognition, RetinaFace, FaceNet, RKNN model conversion