ARFace: Attention-Aware and Regularization for Face Recognition With Reinforcement Learning

Different face regions have different contributions to recognition. Especially in the wild environment, the difference of contributions will be further amplified due to a lot of interference. Based on this, this paper proposes an attention-aware face recognition method based on a deep convolutional neural network and reinforcement learning. The proposed method composes of an Attention-Net and a Feature-net. The Attention-Net is used to select patches in the input face image according to the facial landmarks and trained with reinforcement learning to maximize the recognition accuracy. The Feature-net is used for extracting discriminative embedding features. In addition, a regularization method has also been introduced. The mask of the input layer is also applied to the intermediate feature maps, which is an approximation to train a series of models for different face patches and provide a combined model. Our method achieves satisfactory recognition performance on its application to the public prevailing face verification database.