Methods and systems using depth imaging for training and deploying neural networks for biometric anti-spoofing

Abstract

A method for spoofing detection in a biometric object can include receiving an image pair representing two different viewpoints of the biometric object, and inputting the image pair into a disparity encoder of a trained neural network configured to map the image pair to an encoded signal conveying disparity information associated with the image pair, wherein the disparity encoder was trained for disparity map estimation together with a disparity decoder, the disparity encoder and the disparity decoder forming a disparity branch of the neural network during training. The method can also include inputting the encoded signal into a classification branch of the trained neural network configured to map the encoded signal to a probability that the biometric object is genuine or spoofed, wherein the classification branch was trained for spoofing detection together with the disparity branch. Techniques of training a neural network to perform image-based biometric anti-spoofing are also disclosed.