Authors:
Xin Ning
1
;
Weijun Li
1
;
Meili Wei
2
;
Linjun Sun
1
and
Xiaoli Dong
1
Affiliations:
1
Chinese Academy of Sciences and Cognitive Computing Technology Wave Joint Lab, China
;
2
Cognitive Computing Technology Wave Joint Lab, China
Keyword(s):
Face Anti-spoofing, Convolutional Neural Networks, Stacked Generalized Approach, Intra-class Variations.
Related
Ontology
Subjects/Areas/Topics:
Applications
;
Biomedical Engineering
;
Biomedical Signal Processing
;
Biometrics
;
Biometrics and Pattern Recognition
;
Ensemble Methods
;
Multimedia
;
Multimedia Signal Processing
;
Pattern Recognition
;
Telecommunications
;
Theory and Methods
Abstract:
Thanks for the recent development of Convolutional Neural Networks (CNNs), the performance of face anti-spoofing methods has been improved by extracting more distinguishing features between genuine and fake faces than the hand-crafted texture features. As known, the way of fraud is diverse, thus the fake class has large intra-class variations, so training as a binary classification problem is hard to learn the distinguishing features. In this work, our contribution is a novel model fusion approach for face anti-spoofing, which can reduce the intra-class variations. According to the type of fraud, we firstly train different models for face anti-spoofing problem by CNN, thus the intra-class variations of fake class has reduced during training each model. Distinguishing features can be learned more easily. Then the stacked generalized method is used for combining the lower models to achieve better predictive accuracy. For perfecting the generalized accuracy, the stacked generalized appr
oach changes the weight of each model's prediction, so that the model after fusion can predict precisely whether the face image is fake or genuine. Meanwhile, the experimental results indicate our method can obtain excellent results compared to the state-of-the-art methods.
(More)