Although face recognition has achieved fairly remarkable results in recent years, it heavily relies on large-scale labeled data to train the high-capacity deep convolutional neural networks. It is unrealistic to collect larger labeled datasets to further boost the performance, which requires burdensome and expensive annotation efforts. Meanwhile, there exist numerous unlabeled face images. It is challenging but promising to jointly utilize limited labeled and abundant unlabeled data to obtain higher performance gain, which is the target of semi-supervised learning.
In this paper, we propose a bottom-up method, Adaptive Neighborhood-Aware Attention Network, for semi-supervised face recognition. It includes collaborative relationship prediction based on neighborhood information, and adaptive neighborhood construction based on abnormal neighbor detection. The former is to predict the connection relationship between two nodes, and cluster unlabeled face images based on these relationships, where the neighborhood is defined as a n-hop ego network centered in the given sample called “ego”. Considering the different importance of neighbors, we employ the graph attention network to learn the ego’s representation. The experimental results on MegaFace and IJB-A show that the proposed method can effectively improve the performance of the baseline model.
Moreover, we introduce the abnormal neighbor detection branch to adaptively construct the ego networks of different sizes, which can deal with the imbalanced classes problem. This design reduces the interference of noise in the feature updating process of target samples on the premise that the neighborhood information is rich enough. The experimental results on MegaFace and IJB-A show that the model with the abnormal neighbor detection branch is more robust to noise and yields better performance.
修改评论