| 引用本文: | 程 健1 ,杨高明2 ,杨新露2.基于时空一致性的视频篡改检测方法(J/M/D/N,J:杂志，M：书，D：论文，N：报纸).期刊名称,2025，42（4）：109-115 |
| CHEN X. Adap tive slidingmode contr ol for discrete2ti me multi2inputmulti2 out put systems[ J ]. Aut omatica, 2006, 42(6): 4272-435 |
|
| 摘要: |
| 目的 随着视频篡改技术的快速发展,原始视频与篡改视频之间的差距愈发缩小,现有检测方法在处理篡改
视频数据时,仍需提高检测准确度以及泛化性能,为此,提出一种基于时空一致性的视频篡改检测方法。 方法 首先
通过不同的视频采样步幅预处理视频数据,利用时序卷积核在高采样率视频帧数据中侧重提取帧间时序特征信
息,空间卷积核在低采样率视频数据中侧重提取帧内空间特征信息,并在高采样视频帧数据与低采样视频帧数据
间建立横向连接,从而获得更有效的视频时空特征;同时结合 Transformer 模型在时空特征序列中提取时空特征的
不一致性,实现对篡改视频的判定。 结果 改进的方法在高质量和低质量 FaceForensics++数据集上进行性能测试,
AUC 数值分别达到 99. 47%和 93. 05%,此外在 FaceForensics++数据集上的域内跨伪造方式实验以及 Celeb-DF 数
据集上的跨数据集实验中,测试结果相较于目前主流检测算法同样表现出竞争性,消融实验结果验证了方法中每
个单一模块的有效性。 结论 各项实验结果验证,所提方法在域内性能测试中有着优于现有算法的检测精度,并且
在跨域性能测试中具有更好的泛化性能,即验证了联合时空卷积 Transformer 模型可以提高模型泛化性能。 |
| 关键词: 视频篡改检测 时空一致性 时序卷积核 空间卷积核 Transformer 模型 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Tampering Video Detection Methods Based on Spatiotemporal Consistency |
|
CHENG Jian1 YANG Gaoming2 YANG Xinlu2
|
|
1. School of Artificial Intelligence Anhui University of Science and Technology Anhui Huainan 232001 China
2. School of Computer Science and Engineering Anhui University of Science and Technology Anhui Huainan 232001
China
|
| Abstract: |
| Objective With the rapid development of video tampering technology the gap between original videos and
tampered videos is narrowing. Existing detection methods need to improve accuracy and generalization performance in
detecting video tampering. Therefore a video tampering detection method based on spatiotemporal consistency was
proposed. Methods Firstly the video was processed with different sampling strides. Temporal convolutional kernels were
used to extract temporal features in high-sampling-rate data while spatial convolutional kernels focused on extracting
spatial features in low-sampling-rate data. A lateral connection was established between high-sampling-rate and low-
sampling-rate video data to obtain a better representation of spatiotemporal features. Additionally a Transformer model
was used to extract inconsistencies in the spatiotemporal feature sequence to detect tampered videos. Results The
improved method was tested on the high-quality and low-quality datasets of FaceForensics + + achieving AUC values of
99. 47% and 93. 05% respectively. Furthermore in-domain cross-forgery experiments on the FaceForensics + + dataset
and cross-dataset experiments on the Celeb-DF dataset showed competitive performance compared with current mainstream detection algorithms. Ablation experiments validated the effectiveness of each module. Conclusion Based on the
experimental results of various groups the proposed method demonstrates superior detection accuracy in domain-specific
performance testing compared with existing algorithms. Additionally it exhibits better generalization performance in cross-
domain testing verifying that the combined spatiotemporal convolution and Transformer model can enhance model
generalization performance. |
| Key words: video tampering detection spatiotemporal consistency temporal convolution kernel temporal convolution
kernel Transformer model |
