Ordinal regression with representative feature strengthening for face anti-spoofing

Face anti-spoofing is a crucial link to ensure the security of face recognition. This paper proposes a novel face anti-spoofing method, which performs ordinal regression with representative feature strengthening to learn generalized and discriminative representation for the live and spoof face classification. Specifically, we propose a semantic label schema, which encodes the inter-class ordinal relationships among live and various spoof faces into supervision information to supervise deep neural networks to perform ordinal regression. It enables the learned model to finely constrain the relative distances among features of different categories in the feature space according to the ordinal relationships. The representative feature strengthening network is designed to strengthen important features and meanwhile weaken redundant features for the classification decision. It leverages a dual-task architecture that takes the same single image as input and shares representations via feature fusing blocks. The network first fuses hierarchical paired convolutional features of two streams to learn the common concern of the two related tasks and then, aggregates the learned local convolutional features into a global representation by a learnable feature weighting block. The network is trained to minimize the Kullback–Leibler divergence loss in an end-to-end manner supervised by the semantic labels. Extensive intra-dataset and cross-dataset experiments demonstrate that the proposed method outperforms the state-of-the-art approaches on four widely used face anti-spoofing datasets.