A 3-D Video Watermark Embedding Technology in DCT-CS Domain
Abstract
Digital video watermarking is a critical component of modern information security, particularly for content protection and fingerprinting. This paper presents a novel 3-D embedding scheme that hides digital fingerprint images within the Discrete Cosine Transform - Compressed Sensing (DCT-CS) domain. By combining the data-reduction and encryption properties of Block Compressed Sensing (BCS) with the spatial-temporal robustness of 3-D DCT, the proposed method achieves high transparency and resilience against common video processing attacks.
Introduction
The proliferation of digital video content has necessitated robust mechanisms for copyright protection and unauthorized distribution tracking. Video watermarking designers face a persistent "trilemma" involving three conflicting requirements:
- Robustness: The ability of the watermark to survive signal processing operations such as compression (MPEG-2, H.264), noise addition, and filtering.
- Transparency: Ensuring the watermark is imperceptible to the human visual system (HVS), maintaining the original video quality.
- Capacity: The capability to embed a significant amount of data, such as a high-resolution fingerprint image, without significantly increasing the bit-rate.
Traditional 2-D watermarking methods often fail to account for temporal correlations, making them vulnerable to frame-based attacks like frame dropping or swapping. This research addresses these challenges by leveraging Compressed Sensing (CS) to minimize the watermark payload and 3-D DCT to spread information across both spatial and temporal dimensions.
Methodology
The proposed system follows a structured pipeline for watermark generation, embedding, and subsequent recovery.
Watermark Generation in the CS Domain
Rather than embedding a raw fingerprint image, the system utilizes Block Compressed Sensing (BCS). The original watermark image is divided into blocks and sampled using a pseudo-random measurement matrix. This process serves two purposes: it significantly reduces the data footprint (compression) and provides inherent encryption, as the original image can only be reconstructed if the receiver possesses the correct random seed (the key).
3-D DCT Transformation and QIM Embedding
The video sequence is processed in groups of successive frames. To capture temporal dynamics, a 3-D DCT is applied: first, a 2-D DCT is performed on the spatial planes (x and y axes), followed by a 1-D DCT along the time axis.
The CS-sampled watermark bits are then embedded into the luminance component of I-frames using an adaptive Quantization Index Modulation (QIM) technique. The embedding strength is dynamically adjusted using an empirical Gaussian-based formula to calculate a threshold $T(\alpha)$, ensuring the watermark remains invisible in complex scenes while remaining strong enough for recovery.
Watermark Recovery
At the receiver end, the watermark is extracted from the watermarked video and reconstructed using the Smoothed Projected Landweber (BCS-SPL) algorithm. This iterative method allows for high-quality image recovery from a limited set of CS measurements, provided the measurement matrix key is available.
Experimental Results
The technology was evaluated using standard "Basketball" and "Mobile" video sequences ($720 \times 480$) and a $160 \times 160$ fingerprint image.
- Transparency: The watermarked videos maintained high Peak Signal-to-Noise Ratio (PSNR) values, outperforming benchmark methods and demonstrating superior visual transparency.
- Robustness: The system showed exceptional resilience to compression. Even at bit-rates as low as 0.5 Mbps for MPEG-2 and H.264, the Normalized Correlation (NC) value remained above 0.997.
- Attack Resistance: The watermark remained highly recoverable after various signal processing attacks, including Wiener filtering and Gaussian noise, where competing methods showed significant degradation.
Conclusion
The proposed 3-D DCT-CS watermarking scheme effectively balances the requirements of robustness, transparency, and capacity. By integrating Compressed Sensing with 3-D DCT, the system provides a secure and stable method for protecting digital video content. Future work will focus on extending embedding to P and B frames and exploring advanced information forensics based on CS theory.
Summary updated