Multi-mode multi-layer fusion deep neural network for face anti-spoofing

Abstract

The invention discloses a multi-mode multi-layer fusion deep neural network for face anti-spoofing, and relates to image abnormal sample detection. The deep neural network comprises an image feature extraction front end and a neural network classifier. The network comprises a staggered neural network layer, a multi-modal weight adaptive module and a full connection layer classification unit; wherein the neural network front end comprises a plurality of different modal data processing branches used for respectively processing image data of various different modals, and each branch is formed byconnecting a plurality of residual neural network layers; feature fusion is performed on the image features output by each residual neural network layer of each branch through a multi-modal weight adaptive module; the multi-modal weight self-adaptive module comprises an upper branch and a lower branch, and the upper branch is used for fusing the features of the multi-modal information through an image convolution operation to obtain a fused feature; the lower branch comprises an image convolution operation unit, a global pooling layer, a softmax unit, a ReLU activation unit and a full connection layer.

Description

technical field [0001] The invention relates to image abnormal sample detection, and is a multi-modal multi-layer fusion deep neural network for face anti-spoofing inspection based on an average central loss function. Background technique [0002] As the deep neural network has achieved excellent results in face recognition tasks, face detection technology based on deep learning has been widely deployed in real-world applications, including mobile phone unlocking, access control systems, and face payment. However, face recognition systems based on deep neural networks have been proven to be very susceptible to interference from adversarial examples. These interferences include but are not limited to cracking face recognition systems by printing photos or playing related face videos. Obviously, face authentication through such interference is illegal. Therefore, face anti-spoofing technology has attracted attention. Its main purpose is to detect whether the face captured by ...

AI Technical Summary

Problems solved by technology

[0005] The purpose of the present invention is to solve the problem that in the past, when using deep neural network for face anti-spoofing detection, multi-modal information cannot be used, and the hierarchical information of each layer of neural network cannot be fully utilized. At the same time, due to the single loss function training The model is easy to overfit and other problems, and a multi-modal multi-layer fusion deep neural network for face anti-spoofing based on the average center loss function is proposed

Images