Detection of Double Compression in HEVC Videos Containing B-Frames

Area of Science:

• Digital video compression and processing
• Machine learning for video analysis
• Digital forensics and media integrity

Background:

• Double video compression is a common issue, degrading video quality and posing challenges for forensic analysis.
• Existing detection methods often overlook H.265/HEVC videos with B-frames, a prevalent scenario in modern video streaming.
• Temporal inconsistencies introduced by recompression are key indicators of double compression.

Purpose of the Study:

• To propose and evaluate a novel method for detecting double compression in H.265/HEVC videos, specifically addressing the under-researched area of B-frame inclusion.
• To develop a robust feature extraction and classification framework capable of identifying recompressed video content accurately.
• To assess the practical applicability of the proposed method in realistic double compression scenarios.

Main Methods:

• Extraction of frame-level encoding features, including frame type, coding unit (CU) size, quantization parameter (QP), and prediction modes.
• Representation of each video as a 28-dimensional feature vector.
• Training a bidirectional Long Short-Term Memory (Bi-LSTM) classifier to model temporal inconsistencies arising from recompression.

Main Results:

• The proposed method achieved a detection accuracy of 80.06% on a custom dataset of 129 HEVC-encoded videos.
• The approach demonstrated superior performance compared to two existing baseline methods.
• The results indicate the method's effectiveness in identifying double compression in realistic scenarios.

Conclusions:

• The developed method provides an effective solution for detecting double compression in H.265/HEVC videos, even those containing B-frames.
• The use of Bi-LSTM classifiers trained on specific encoding features shows promise for video forensic applications.
• The study highlights the importance of addressing B-frame scenarios for comprehensive double compression detection.