Graph-Based Fingerprinting for Robust Video Sequence Identification under Temporal Reordering

Abstract

Video fingerprinting plays a vital role in near-duplicate detection, copyright protection, and multimedia retrieval. Most existing fingerprinting techniques rely on frame-level or sequential representations and assume that the temporal order of video content remains intact. However, real-world video reuse often involves temporal distortions such as segment reordering, frame dropping, and partial clip extraction, which significantly degrade the performance of sequence-dependent methods. This paper proposes a Graph-Based Video Fingerprinting Framework (GBVFF) for robust video identification under temporal distortions such as reordering, frame dropping, and segment shuffling. Unlike conventional sequence-dependent approaches, the proposed method models videos as graph structures to preserve contextual relationships independent of temporal order. The framework integrates Mean Canberra Distance for similarity estimation, and KL-divergence-based selection of representative fingerprint features. Experimental evaluation on benchmark datasets, including YouTube-8M and WebVid, demonstrates that GBVFF achieves a 12.4% improvement in accuracy, 14.8% higher precision, and a 27.6% reduction in false positive rate compared to state-of-the-art methods. The results validate that graph-based representations significantly enhance robustness against temporal perturbations, making the approach effective for real-world video retrieval, duplicate detection, and copyright protection systems.