Face recognition system with feature normalization

ABSTRACT\r\n\r\nThis study proposed an identity verification system that uses face recognition. The system features face detection as well as facial feature extraction and comparison methods. Early methods of face detection involved using specific approaches coupled with a classifier to extract features and detect faces. Although these methods can detect faces quickly, they generate a high false positive rate. Recent face detection methods based on a deep learning structure are extremely accurate but time-consuming. This study realized a face detection method based on the histogram of oriented gradient. The proposed method is not as accurate as deep learning; however, it is fast and can complete instant computing. Early methods of face recognition also involved using feature extraction methods coupled with a classifier to complete face recognition; however, these methods were not extremely accurate. The emergence of deep learning has facilitated greatly increasing the accuracy of face recognition. A deep learning-based method requires the entire deep learning structure to be retrained when a system needs to add a new user. This requirement is not feasible in actual applications. A researcher has therefore proposed the FaceNet method, which uses deep learning structure to extract eigenvectors and calculates the distance between eigenvectors as a measure of face similarity. Thus, the entire deep learning structure does not need to be retrained when a new user is added to the system. In this paper, FaceNet was used to extract the eigenvector of a face. However, the experiments of this study showed that facial features extracted using FaceNet are unevenly distributed in different dimensions, and using the calculated distance of the eigenvector as a measure of face similarity will yield inaccurate results. Therefore, this study proposed a facial feature normalization comparison method. The experimental results verified that the proposed method can achieve more than 98% accuracy and can be applied in practice.