摘要
针对目前人脸伪造检测中提取的面部特征不充分、检测准确率低等问题,提出了一种基于空域频域相结合的深度伪造检测方法CSFNet,由空域、频域双流网络组成,空域分支利用SRM滤波器和注意力模块捕捉真实与篡改区域间的差异,频域分支使用4个可学习的频率滤波器挖掘伪造模式。结果表明:CSFNet在压缩和跨数据集情况下检测的准确率均有不同程度的提升,具有一定的迁移性。此外,消融实验验证了每个模块的有效性。
To address the problems of inadequate extracted facial features and low detection accuracy in current face forgery detection,a deep forgery detection method CSFNet based on a combination of space domain and frequency domain is proposed,consisting of a dual-flow network in the space domain and frequency domain,with the space domain branch using SRM filters and attention modules to capture the differences between real and tampered regions,and the frequency domain branch using four learnable frequency filters to mine forgery patterns.The results show that the detection accuracy of CSFNet is improved in different degrees under compression and cross-data sets,and it has certain migration.In addition,ablation experiments verified the effectiveness of each module.
作者
程晴晴
范智贤
CHENG Qingqing;FAN Zhixian(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan Anhui 232001,China)
出处
《兰州工业学院学报》
2023年第4期91-95,共5页
Journal of Lanzhou Institute of Technology
关键词
人脸伪造检测
深度伪造
频域
双流网络
SRM滤波器
Face forgery detection
depth forgery
frequency domain
dual-stream network
SRM filters
