Face anti-fraud method based on cross-domain feature alignment network

Abstract

The invention provides a face anti-fraud method based on a cross-domain feature alignment network, and belongs to the technical field of mode recognition. The method comprises the following steps: step 1, taking ResNet18 as a basic feature generation model, and designing an improved feature generation network in order to relieve domain difference and enhance deep feature representation; and step 2, in order to explore multi-granularity feature alignment to improve generalization ability for unknown target scenes, the invention provides a multi-granularity feature alignment network to perform local region and global image feature alignment. And 3, in order to reduce the intra-class distance and increase the inter-class distance at the same time, total loss is adopted to calculate the classification loss of the network, and a final network model is obtained through iterative adversarial training and parameter updating. Under the condition that a small amount of calculation is increased, the domain difference is effectively relieved, meanwhile, the feature expression ability is enhanced, a more robust and generalized feature space is captured, and a clearer classification boundary is achieved.

Description

technical field [0001] The invention belongs to the technical field of computer pattern recognition, in particular to a face anti-fraud method. Background technique [0002] In recent years, the rise of deep learning has promoted the long-term development of face recognition technology, and more and more face recognition systems have been deployed in various application scenarios, such as fast payment, public security, and border control. However, the emergence of various representation attacks, such as image printing attacks, video replay attacks, and 3D mask attacks, makes the current face recognition system unreliable. Therefore, face anti-spoofing is one of the face recognition systems The extremely critical component has become a research hotspot in academia and industry, involving multiple interdisciplinary research, of which biology, communication, and computer science are the main disciplines, and it is a relatively new and promising research direction. [0003] At ...

AI Technical Summary

Problems solved by technology

However, methods based on domain adaptation have the following drawbacks: 1) Since domain adaptation methods require unlabeled target domain samples to participate in the training process, in many practical scenarios, it is difficult and expensive to collect a large amount of unlabeled target data for training, Not even information about the target domain is available

The methods based on domain generalization can well solve the defects of domain adaptive methods, but the methods based on domain generalization have the following deficiencies: 1) These methods tend to simply consider directly aligning the entire feature space, therefore, their essence is Coarse-grained domain alignment method, which can cause performance bottlenecks

2) Furthermore, most of these methods only focus on the design of domain-aligned architectures, while paying little attention to the impact of feature generation networks on the final generalization performance

Images