摘要
针对虚假视频的检测中,特征噪声多、数据量大和检测准确率低的问题,提出一种改进高斯滤波对地标点进行去噪和改进深度网络模型来提升精度的算法(IGFNet);算法将高斯滤波拓展到空域和值域上,在滤除低频噪声的同时尽可能保留高频噪声,使得在后续特征数据的处理中,标记点的精度得以提升;采用特征化的数据代替原始数据以减少数据量,使得送入网络的数据量大大缩小,有效缩短了训练时长和减少了网络参数量,在实际应用中可以增加鉴别虚假视频的效率;并且针对人脸属性特征点的差异采用不同深度的双流神经网络,从而更加有效地学习如何鉴别虚假视频。实验表明:改进高斯滤波网络算法(IGFNet)有效地增加了真假脸检测的准确率,在与当前较为优秀的方法如Meso4,Xecption,LRNet等的对比下,IGFNet的准确率均有不同程度的提升,在跨数据集的测试上提升尤为明显;在压缩过的虚假视频测试中,IGFNet的泛化下降程度最低,显示了较强的鲁棒能;通过改进网络加入梯度热力图以直观判断出IGFNet对于深度伪造图片的鉴别能力。
Aiming at the problems of excessive feature noise,a large amount of data and low detection accuracy in the detection of fake videos,an algorithm was proposed to improve the Gaussian filter to denoise the landmark points and improve the deep network model,so as to improve the accuracy.The Gaussian filter is extended to the spatial domain and the value domain,and the high-frequency noise is retained as much as possible while filtering out the low-frequency noise so that the accuracy of the marker points can be improved in the subsequent processing of the feature data.The characteristic data is used instead of the original data to reduce the amount of data,which greatly reduces the amount of data sent to the network,effectively shortens the training time and reduces the number of network parameters,and can quickly identify fake videos in practical applications.In addition,according to the difference of face attribute feature points,dual-flow neural networks with different depths are used to learn how to identify fake videos more effectively.The experimental results show that the improved Gaussian Filter network algorithm(IGFNet)effectively increases the accuracy of real and fake face detection.Compared with the current excellent methods such as Meso4,Xecption and LRNet,the accuracy of IGFNet has been improved to different degrees,especially in the test across datasets.In the test of compressed fake videos,IGFNet has the lowest degree of generalization degradation,showing strong robustness.The ability of IGFNet to discriminate deep forgery images is intuitively judged by improving the network and adding the gradient heat map.
作者
瞿远近
吴起
QU Yuanjin;WU Qi(School of Computer Science and Engineering,Anhui University of Science and Technology,Anhui Huainan 232001,China;School of Artificial Intelligence,Anhui University of Science and Technology,Anhui Huainan 232001,China)
出处
《重庆工商大学学报(自然科学版)》
2023年第4期41-47,共7页
Journal of Chongqing Technology and Business University:Natural Science Edition
基金
安徽省自然科学基金项目(2008085MF220)
安徽省高校自然科学基金项目(KJ2019A0109)
安徽省重大科技专项基金项目(18030901025)
安徽理工大学研究生创新基金项目(2021CX2102).
关键词
深度学习
深度图像伪造
改进高斯滤波
循环神经网络
deep learning
deep image forgery
improved Gaussian filtering
recurrent neural networks