Face Recognition Speed Optimization Method for Embedded Environment

To solve the problem that it is difficult to combine performance and speed of face recognition in embedded environment, we design and implement a fast face recognition method based on deep learning. Firstly, face detection and quality selection are performed based on MTCNN algorithm to reduce the amount of feature extraction data needed for recognition. Secondly, the RKNN model is used to quantify the acceleration of feature extraction. Finally, fast face recognition is realized by feature matching. The experiment shows that the method is beneficial for face recognition processing on the RK3399Pro platform, and the average time is shortened by 80% compared with that before acceleration. In the case of ensuring the feature extraction accuracy and recognition accuracy, it basically meets the real-time requirements.