Novel Real-time Face Recognition from Video Streams

Real-time recognition from video streams is a very challenging task due to background, facial expressions, and lighting differences. Recent studies show that deep learning approaches can achieve impressive performance on this task. Our system solves these problems well using deep learning method. It contains face detection module and face recognition module. Our face detection module is based on MTCNN, which is very fast and accurate, and is robust enough for changes in lighting differences and background. Our face recognition module is based on FaceNet, which directly learns mapping the face images to the points in Euclidean space, where the distance of two points in Euclidean space directly corresponds to how similar two face images are. Once such Euclidean space is created, we transform the face images into the FaceNet embedding, as the feature vectors for face images. Next, we put the feature vectors into an SVM classifier to help us quickly identify the face images. The experimental results show that the system has very high accuracy and low computational complexity, which is the key to real-time face recognition, and has significant value in application.