Robust Face Anti-spoofing Model With New Convolutional Computation CDC
Three main points
✔️ Introducing a new convolutional operation called CDC, CDCN is proposed, extracting more invariant and unique features than before, and building a more robust Face Anti-spoofing(FAS) model
✔️ CDCN++ is built, extending CDCN First approach to search the CDC's search space for face anti-spoofing architecture by NAS (Network Architecture Search)
✔️ CDC-based model achieves SOTA for unknown spoofing methods and shows high robustness
Searching Central Difference Convolutional Networks for Face Anti-Spoofing
written by Zitong Yu, Chenxu Zhao, Zezheng Wang, Yunxiao Qin, Zhuo Su, Xiaobai Li, Feng Zhou, Guoying Zhao
(Submitted on 9 Mar 2020)
Comments: Accepted at CVPR2020
Subjects: Computer Vision and Pattern Recognition (cs.CV)
Paper Code
overview
By 2020, various demonstration tests will be launched in various locations, including entrance gates and vending machines using face recognition, and the introduction of face recognition is progressing rapidly. To manage staff personnel. Facial recognition will be introduced. Various advantages of biometrics, such as high security, hands-free and contactless, are attracting attention.
On the other hand, however, there are also concerns about Presentation Attacks. It is possible to trick facial recognition systems by simply presenting a printed image or video. There have also been reports of spoofing using 3D masks.
Source: fig.1, 3D Mask Face Anti-Spoofing, P C Yuen et al., Department of Computer Science, Hong Kong Baptist University
To make facial recognition systems more trustworthy, technology to detect such spoofing (Presentation Attack Detection, PAD) is essential.
So far, two methods have been proposed in the field of PAD: one based on handcrafted features and the other based on deep learning features. Handcrafted methods (ex. Local Binary Pattern, LBP) use local information as features to represent invariant details (ex. color texture, moiré patterns, noise artifacts) of real and spoofed details, and are often robust , Convolutional Neural Networks (CNNs) have been reported to have a high representational capability to distinguish real from spoofing, often through a series of convolutional operations with nonlinear activation.
However, CNN-based methods, due to their categorization capabilities, tend to be difficult to represent the details between real and spoofed images, and tend to be greatly influenced by the external environment, such as differences in lighting and cameras.
Therefore, in this paper, we propose a new convolutional operation called Central Difference Convolution (CDC), which overcomes the weaknesses by integrating local information into the CNN and provides robust feature representation that is robust to the external environment. We also use Neural Architecture Search (NAS) in a specially designed CDC search space to build an optimal network for face anti-spoofing.
To read more,
Please register with AI-SCHOLAR.
ORCategories related to this article
