If you google "face anti-spoofing" or "liveness detection," you will see many articles and research papers which try to accomplish this task. There also looks to be a lot of open source code you can try out online
Face anti-spoofing/liveness detection can be thought of as a binary classification task (i.e. does a given image or video stream contain a genuine face or not). This isn't my research area, but I would assume there are two main routes you can take
Motion approach: make user blink or move in a way which convinces you that they are real. This will likely involve ML models to detected the desired action in a video stream
Feature approach: extract useful features from an image/video stream frames (might require the camera captures depth information in addition to color) and use the features to make a classification decision. This will almost certainly involve ML models to extract features and perform the final classification
There are probably hybrid approaches as well. Based on your static image constraint, (2) is probably where you want to start
The approaches that you've mentioned mostly are used for video based biometrics. There's a segment of face anti spoofing techniques known as print attack where an image of the face is used to trick the biometric. But i couldn't find any model/dataset specifically for a case where the face is compared to a static image. That's what I had researched so far, but I'll still go through with your recommendations and try to find something.
Based on your description, something like (2) is what you're going to need. If you deal with static images you are going to need to extract some features (maybe using ML or traditional CV technique) then use a binary classifier to determine if it is a valid facial image or one involving spoofing (including the print attack you mention)
Furthermore, I disagree. (2) is not really a video-based approach. Please look closely at the tutorial and consider the logic of what he is doing. He is taking each frame in the training videos as static images. Then he uses the resulting images to train an ML model. You could substitute those video frame images with any images that fit your preference, and the overall workflow would be the same
Asked: 2020-03-17 02:16:04 -0600
Seen: 685 times
Last updated: Mar 17 '20
