关键词: 视频篡改检测, 目标移除, 可学习Ptuning, 多尺度特征交互, 多视图特征

Abstract: With the continuous advancement of artificial intelligence and big data technologies, the threshold for making fake videos has been significantly reduced. Therefore, identifying whether a video has been tampered with is crucial for ensuring the authenticity and credibility of the information. Current mainstream video forgery detection methods rely on convolutional neural networks, which exhibit limited capability in capturing temporal dependencies and lack comprehensive understanding of global temporal patterns. To address this issue, this paper proposes a learnable Ptuning based method for video object removal tamper detection and localization. Firstly, the prior knowledge of the pretrained model is fully mined by learnable Ptuning, and multiview features such as spatial, temporal and highfrequency are efficiently extracted. Secondly, a multiscale feature interaction module is proposed to accurately capture the tampering traces from finegrained to coarsegrained through multiscale convolution operation and twostep decomposition strategy. Furthermore, a multiview fusion attention module is designed to significantly enhance the information sharing and fusion ability among multiview features via the crossview interaction mechanism. Experimental results demonstrate that the proposed method outperforms existing detection methods in both the time domain and the spatial domain.

Key words: video tamper detection, object removal, learnable Ptuning, multiscale feature interaction, multiple view feature