Face Quality Assessment via Semi-supervised Learning

Face Quality Assessment via Semi-supervised Learning

연구 분야: Artificial Intelligence

학회: ICCPR '19: Proceedings of the 2019 8th International Conference on Computing and Pattern Recognition

Face quality assessment, used for selecting a "good" subset from face images captured over multiple frames in uncontrolled conditions, plays a significant role in video-based face recognition. By removing the poor quality images, it can not only improve recognition performance but also reduce the computation cost. This paper proposes an end-to-end face quality assessment algorithm based on a semi-supervised learning framework. The contributions of the proposed method are threefold. (i) Making use of unlabeled data from target domain to fine-tune a neural network by a strategy of automatically updating labels. (ii) Combining prior knowledge with feature learning by using a set of characteristics as binary constraints. (iii) Proposing a light neural network model for training and predicting. Experiments demonstrate that our model can get much higher accuracy in face quality assessment task than the models trained with the same amount of labeled faces, meanwhile the complexity is lower. Experimental results also show that our method can improve the performance of face recognition by face selection.