作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2022, Vol. 48 ›› Issue (2): 194-200,206. doi: 10.19678/j.issn.1000-3428.0060267

• 图形图像处理 • 上一篇    下一篇

基于自注意力胶囊网络的伪造人脸检测方法

李柯, 李邵梅, 吉立新, 刘硕   

  1. 解放军战略支援部队信息工程大学 信息技术研究所, 郑州 450003
  • 收稿日期:2020-12-14 修回日期:2021-01-20 发布日期:2021-01-25
  • 作者简介:李柯(1995-),男,硕士研究生,主研方向为图像处理、深度学习、计算机视觉;李邵梅,副教授;吉立新,研究员、博士生导师;刘硕,硕士研究生。
  • 基金资助:
    国家自然科学基金青年科学基金项目(62002384)。

Method of Face Forgery Detection Based on Self-Attention Capsule Network

LI Ke, LI Shaomei, JI Lixin, LIU Shuo   

  1. Institute of Information Technology, People's Liberation Army Strategic Support Force Information Engineering University, Zhengzhou 450003, China
  • Received:2020-12-14 Revised:2021-01-20 Published:2021-01-25

摘要: 当前以换脸为代表的伪造视频泛滥,给国家、社会和个人带来潜在威胁,有效检测该类视频对保护个人隐私和维护国家安全具有重要意义。为提高视频伪造人脸检测效果,基于可解释性好的胶囊网络,以Capsule-Forensics检测算法为基础,提出一种结合自注意力胶囊网络的伪造人脸检测方法。使用部分Xception网络作为特征提取部分,降低模型的参数量,在主体部分引入带注意力机制的胶囊结构,使模型聚焦人脸区域,将综合多维度的Focal Loss作为损失函数,提高模型对难分样例的检测效果。实验结果表明,与Capsule-Forensics算法相比,该方法能够减少模型参数量和计算量,在多种伪造类型数据集上均具有较高的准确率。

关键词: 伪造人脸检测, 胶囊网络, 模型可视化, 注意力机制, 损失函数

Abstract: In recent years, face forgery is abused in fake videos, imposing a potential threat on the national, social and individual level, so face forgery detection is of great significance to individual privacy protection and national security.To improve the performance of face forgery detection for fake videos, a face forgery detection method that combines a self-attention capsule network with the Capsule-Forensics algorithm is proposed.This method uses part of the Xception network for feature extraction, which reduces the number of parameters.Then a capsule structure with attention mechanism is introduced into the main part to make the model focus on the facial area.Finally, the comprehensive multi-dimensional Focal Loss is used as the loss function to improve the detection effect of the model for indistinguishable samples.Experimental results show that compared with the Capsule-Forensics algorithm, this method can reduce the number of model parameters and the amount of computation, while displaying a higher accuracy on multiple forgery data sets.

Key words: face forgery detection, capsule network, model visualization, attention mechanism, loss function

中图分类号: