Opportunity
The new Versatile Video Coding (H.266/VVC) standard offers improved compression but introduces complex new coding tools (multi-type tree partitioning, adaptive loop filtering) that interfere with existing video forensics methods. Detecting whether a video has been recompressed multiple times—even with the same codec and parameters—is critical for verifying authenticity and spotting manipulation. Traditional detection methods fail under same-parameter recompression because traces become extremely faint. There is a need for a forensic method specifically designed for H.266/VVC that can reliably detect double/multiple compression when encoding parameters are identical, which is a more challenging and realistic scenario.
Technology
This patent presents a system with a recompression detection engine (machine-learning based) that analyzes compressed video to determine if it has been processed multiple times using the same H.266/VVC codec and identical coding parameters. The method leverages two novel features extracted from I-frames during decoding.
First, it calculates the Variation of Coding Unit (CU) Partition Modes. Since recompression causes CU partition instability, the system performs Minimum Unit Mapping (MUM) to divide all CUs into 4×4 minimum units, labeling each by its parent CU type. It then measures the change in the number of these units for each label across consecutive compressions.
Second, it calculates the Consistency of Prediction Modes. Using Subunit Prediction Mapping (SPM), it maps intra prediction modes down to 4×4 subunits, then computes consistency ratios of adjacent prediction mode pairs in horizontal, vertical, and diagonal directions across compressions.
These two feature sets (48-D variation + 12-D consistency) are fused into a 60-D feature vector and fed into an SVM classifier. The engine is trained on positive samples (recompressed twice with same parameters) and negative samples (compressed once only). Extensive experiments across CQP and ABR rate control modes, different GOP sizes, and 4K resolutions validate the approach.
Advantages
- Detects Same-Parameter Recompression: Specifically addresses the hard problem of detecting double compression when encoding parameters are identical, where most methods fail.
- High Accuracy: Achieves 97.33% average detection accuracy across various QP values, outperforming reference methods by 3-5%.
- Robust to Coding Modes: Works effectively under both Constant QP (CQP) and Average Bitrate (ABR) rate control modes (96.79% ABR accuracy).
- Resilient to GOP Variations: Maintains high accuracy (88-99%) even as Group of Pictures size increases, though performance depends on I-frame frequency.
- Fusion Feature Advantage: The fused 60-D feature consistently outperforms either sub-feature alone.
- Validated on 4K Video: Performs well on high-resolution UHD content from the SJTU-4K dataset.
Applications
- Video Forensics & Authenticity Verification: Detecting whether a video has been re-edited and re-saved (recompressed) to claim originality.
- Digital Evidence Integrity: Validating that surveillance or legal video evidence has not been manipulated after capture.
- Copyright Infringement Detection: Identifying videos that have been recompressed to remove watermarks or alter metadata.
- Social Media Content Analysis: Flagging re-uploaded, manipulated, or repurposed video content on platforms.
- Security & Tamper Detection: Detecting deepfake or spliced videos where multiple compressions may indicate forgery attempts.
