Efficient and Real-Time Face Recognition Based on Single Shot Multibox Detector

Abstract

In this paper we present an efficient and real-time human face detection and recognition method based on human body region of interests (ROIs) provided by the single shot multibox detector (SSD). The SSD is a state-of-art general purpose object detector that can detect all kinds of items in the image data and provides detection probabilities. On the other hand, the histogram of oriented gradients (HOG) is another superb detector that is specially designed for human face detection. However, it takes much time to scan the whole image data in order to get the face features. Hence, the issue to us is to reduce the computation time spent for searching human faces and to cope with scalability of the object sizes. Here, in our method, we place the SSD in front of the HOG. The SSD is used to make the ROIs of the human bodies, not the faces importantly, so that the image data containing the human body ROIs only are processed by the HOG. In this way, the HOG can save much time to produce the ROIs of human faces. Then, the feature vectors for the human face ROIs are computed in order to train and also to recognize the people’s identities by using a deep learner. The computer simulations are performed to verify the proposed system using several well-known data bases. The performance evaluation is done in terms of speedup and accuracy as the multiplicity and scalability of people changes. The results show us that the proposed system performs efficiently and robustly than that of the conventional system without SSD, and advantageously it comes with better real-time feasibility.