3D Convolutions and Complex Wavelets Strengthen Invisible Video Watermarks

Digital video now carries much of the world’s copyrighted and security-sensitive material, and embedding an invisible watermark that survives both viewing and compression remains one of multimedia engineering’s persistent challenges. A study published in Scientific Reports by Dantong Li and Guixin Li proposes a video watermarking algorithm that combines three-dimensional convolutions with the dual-tree complex wavelet transform, or DT-CWT, reporting measurable gains in both robustness and visual quality on a public benchmark.

The core distinction lies in how the wavelet transform is used. In earlier convolutional neural network watermarking methods built around DT-CWT, the transform served mainly as a static frequency-domain preprocessing step. The new algorithm instead models DT-CWT sub-band features across frames and through spacetime using a three-dimensional convolutional network, allowing the embedding process to account for how frequency-domain information evolves from one video frame to the next rather than treating each frame in isolation.

Two mechanisms drive the adaptive placement of the watermark. Dense feature fusion combines information from multiple network layers, while an attention guidance mechanism directs the system toward the locations and frequency sub-bands best suited to carry hidden data. Together, the authors report, these components achieve watermark embedding that adapts both in the frequency domain and along the temporal consistency dimension, so the mark remains coherent as the video plays.

Compression is treated as a first-class problem rather than an afterthought. Real-world video is almost always encoded with standards such as H.264, and the distortion introduced during encoding can degrade a watermark. During training, the algorithm employs a differentiable compression simulation of H.264 encoding, which lets the network perceive the actual distribution of encoding distortion. According to the study, this improves the stability of the watermark in real compression environments because the embedding learns to tolerate the specific losses the codec will impose.

The experimental evaluation used the publicly available UVG dataset, a standard collection of high-quality video sequences. In ablation experiments, which remove individual modules to test their contribution, the complete model achieved a peak signal-to-noise ratio of 41.8 dB and a structural similarity index of 0.985, results the authors cite as validating the importance of the key components.

Robustness was assessed against a battery of typical attacks. After 100 training epochs, the proposed algorithm reached a bit error rate of 0.22 on the test set while maintaining a recognition rate above 96 percent under various typical attacks. The method proved more resistant to scaling, cropping, rotation, and Gaussian noise than the comparison approaches, meaning the hidden watermark could still be recovered after the video had been geometrically altered or degraded by noise.

Quality comparisons against existing methods showed consistent advantages across several indicators. The peak signal-to-noise ratio of the new algorithm was improved by an average of about 2 to 4 dB relative to the comparison method, and the structural similarity index rose by roughly 0.03 to 0.06. The video multimethod evaluation fusion index, a perceptual quality metric widely used in streaming pipelines, improved by about 6 to 12 points, indicating that the watermarked videos remained closer to the visual fidelity of the originals.

The authors position the algorithm as a high-performance practical solution that significantly enhances the robustness and stability of watermarks while preserving video quality. Because the results are reported on a single public dataset and compared against specific baseline methods, the findings describe performance within that experimental scope; the ablation and comparative figures are drawn from the study’s own test conditions rather than from independent deployment. Even so, the reported combination of spacetime-aware wavelet modeling, attention-guided embedding, and codec-aware training offers a concrete technical template for watermarking systems that must survive the compression and manipulation that digital video routinely undergoes in distribution and storage.

Subject of Research: A deep learning video watermarking algorithm integrating 3D convolutions with the dual-tree complex wavelet transform for robust, imperceptible copyright protection

Article Title: Digital video watermarking algorithm integrating 3D convolution and DT-CWT

Article References: Li, D., & Li, G. (2026). Digital video watermarking algorithm integrating 3D convolution and DT-CWT. Scientific Reports. https://doi.org/10.1038/s41598-026-76098-4

Image Credits: AI Generated

DOI: 10.1038/s41598-026-76098-4

Keywords: video watermarking, 3D convolution, dual-tree complex wavelet transform, H.264 encoding, digital rights protection, multimedia security, peak signal-to-noise ratio, structural similarity index, UVG dataset, attention mechanism, robustness, Scientific Reports