Scholarlink

<p>The variety of cameras, frame rates, and motion patterns in real-world videos makes it difficult to detect frame-level tampering. This study presents a unified, training-free, three-stage framework to detect, localize, and classify inter-frame forgery in digital videos. Instead of using fixed thresholds, the proposed framework employs frames-per-second (FPS)-adaptive thresholding, enabling reliable operation across mobile, CCTV, and high-FPS videos. Stage 1 uses curl-based motion magnitude and a coefficient-of-variation model to identify anomalies; Stage 2 employs z-score modeling for block-wise rotational motion analysis; and Stage 3 uses geometric moments and sliding-window pattern alignment to classify different types of forgeries. The framework achieves average detection accuracy, precision, recall, F1-score, and localization accuracy of 97.43%, 97.48%, 99.09%, 98.27%, and 96.73%, respectively, across five datasets comprising three public datasets and two datasets generated using the proposed forgery-generation framework, demonstrating consistent cross-dataset performance. In addition to providing a reliable, interpretable, and computationally efficient forensic framework, comparative analysis shows how the proposed framework addresses important research gaps.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

VortiFlow: curl-based block-wise motion dynamics for frame-level video forgery detection, localization, and forgery type classification

  • 期刊论文

  • 发表日期：

    2026年8月3日

查看全文
  • Manpreet Kaur Aulakh,
  • Navdeep Kanwal,
  • Manish Bansal

摘要

The variety of cameras, frame rates, and motion patterns in real-world videos makes it difficult to detect frame-level tampering. This study presents a unified, training-free, three-stage framework to detect, localize, and classify inter-frame forgery in digital videos. Instead of using fixed thresholds, the proposed framework employs frames-per-second (FPS)-adaptive thresholding, enabling reliable operation across mobile, CCTV, and high-FPS videos. Stage 1 uses curl-based motion magnitude and a coefficient-of-variation model to identify anomalies; Stage 2 employs z-score modeling for block-wise rotational motion analysis; and Stage 3 uses geometric moments and sliding-window pattern alignment to classify different types of forgeries. The framework achieves average detection accuracy, precision, recall, F1-score, and localization accuracy of 97.43%, 97.48%, 99.09%, 98.27%, and 96.73%, respectively, across five datasets comprising three public datasets and two datasets generated using the proposed forgery-generation framework, demonstrating consistent cross-dataset performance. In addition to providing a reliable, interpretable, and computationally efficient forensic framework, comparative analysis shows how the proposed framework addresses important research gaps.