Design and Implementation of an Asymmetric Face Recognition System Based on Unsupervised Learning

Abstract This paper mainly discusses the asymmetric face recognition problem where the number of names in a name list and the number of faces in the photo might not be equal, but each face should be automatically labeled with a name. The motivation for this issue is that there had been many meetings in the past. After each meeting, the participant took group photos. The meeting provided only a corresponding name list of participants without one-to-one labels. In the worst case, the group photo might mix with the faces that were not participating in the meeting. Another reason for asymmetric face recognition is that some meeting personnel did not appear in photos because they assisted in taking pictures. This paper proposes an Asymmetric Face Recognition Mechanism, called AFRM in short. Initially, the proposed AFRM adopts the Histogram of Oriented Gradients (HOG) and Support Vector Machine (SVM) to detect and extract all faces from photos. Next, AFRM extracts the features from each face using the convolution feature map (Conv_FF)and adopts the features to partition the faces into different classes. Then the AFRM applies the statistic-based mechanism to map each name in the name list to each face class. According to this mapping, each face will be associated with one name. To quickly identify a face during the meeting, the AFRM applies the K-Nearest Neighbors (KNN) to represent the features of each face. During the new meeting, the proposed AFRM can extract the feature of one face and then adopts KNN to derive the features. Experimental results show that the proposed mechanism achieves more than 97% accuracy without one-to-one name and face labeling.