《计算机应用》唯一官方网站

• •    下一篇

基于区域分块和轻量级网络的人脸反欺骗模型

贺丹1,何希平1,2,李悦1,袁锐1,牛园园1   

  1. 1.重庆工商大学 人工智能学院,重庆 400067;
    2.检测控制集成系统重庆市工程实验室(重庆工商大学),重庆 400067

  • 收稿日期:2021-10-08 修回日期:2021-12-14 接受日期:2021-12-23 发布日期:2021-12-31 出版日期:2021-12-31
  • 通讯作者: 何希平
  • 基金资助:
    重庆市研究生科研创新项目;重庆工商大学研究生科研创新项目

Face Anti-spoofing Model Based on Regional Blocks and Lightweight Network

  • Received:2021-10-08 Revised:2021-12-14 Accepted:2021-12-23 Online:2021-12-31 Published:2021-12-31
  • Supported by:
    Scientific Research and Innovation Foundation of Chongqing;Graduate Innovation Project of Chongqing Technology and Business University

摘要: 如何高效的辨别各种被攻击的人脸是人脸识别过程中迫切需要解决的问题。基于深度学习的人脸反欺骗算法在有着高性能的同时,也带来了庞大的参数量和计算量,使其无法部署在移动设备中。针对以上问题,提出了一种基于区域分块和轻量级网络的人脸反欺骗模型。首先,对训练样本进行随机区域分块;然后,设计了一种基于注意力机制的轻量级网络用于特征提取和分类;最后,为了提高测试准确度,对测试样本进行基于区域分块的数据扩增。实验结果表明,该模型在CASIA-FASD和REPLAY-ATTACKR数据集上达到了100%的准确率,是目前效果最好的算法;而在CASIA-SURF数据集的Depth模态上也获得99.49%的准确率和0.4580%的平均分类错误率(ACER),远优于ResNet、ShuffleNet等卷积神经网络;模型的参数量也仅有0.2582MB。在实际应用中,端到端的轻量级网络结构更方便部署在移动设备上进行实时的人脸反欺骗检测。

关键词: 人脸反欺骗, 区域分块, 中心差分卷积, 注意力机制, 轻量级网络, 卷积神经网络

Abstract: How to effectively identify all kinds of spoofed faces is an urgent problem to be solved in the process of face recognition. While the face anti-spoofing algorithm based on deep learning has high performance, it also brings a large number of parameters and calculation, so it can not be deployed in mobile or embedded devices. To solve the above problems, a face anti-spoofing model based on regional block and lightweight networks was proposed. Firstly, the training samples were randomly cut into some patches. Then, a lightweight network based on attention mechanism was designed for feature extraction and image classification. Finally, in order to improve the detection performance, data augmentation was conducted on the test samples based on the method of regional block. The experimental results show that the proposed algorithm achieves 100% accuracy on the datasets REPLAY-ATTACK and CASIA-FASD, which achieves the state-of-the-art. 99.49% accuracy and 0.4580% Average Classification Error Rate (ACER) are also obtained on the Depth modal of the CASIA-SURF dataset, which are much better than that obtained with other convolutional neural networks such as ResNet, ShuffleNet. The parameter amount of the model is only 0.2582MB. In practical applications, the end-to-end lightweight network architecture makes it easier to deploy on mobile devices for real-time face anti-spoofing detection.

Key words: face anti-spoofing, regional block, central difference convolution, attention mechanism, lightweight network, convolutional neural network

中图分类号: