Exploring Hybrid Pooling of Pretrained Residual Network for Face Anti-spoofing
Daoxiang Zhou, Yunfei Zhang, Jun Zhang · 2023
Face anti-spoofing has become an important part of face recognition system, because it can detect non-living faces in advance, thus the security of the subsequent identification process can be ensured. According to the literature analysis and statistics, the discrimination of handcrafted features is unsatisfying and the high computation burden and overfitting issue of deep learning methods are still unsolved. In this paper, an effective and simple feature extraction method for face anti-spoofing is proposed, which consists of three parts. Firstly, for a given face sample, it is fed to a pretrained residual network, the feature map of a certain layer is extracted. Secondly, first- and second-order pooling are performed simultaneously on the feature map, the goal of which is to learn more discriminative features. Thirdly, the linear support vector machine is exploited to distinguish genuine faces from spoofing ones. The feature distribution visualization and decision scores density show that our learned features possess good intra-class similarity and inter-class diversity. Extensive experiments are conducted on two widely-used spoofing datasets in terms of three standard evaluation metrics: equal error rate, half-total error rate, and area under curve. The comparison results can demonstrate the effectiveness of our proposed method evidently.