Video Watermark Encoding With Imperceptible Spatial-Temporal Embedding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video watermarking technologies often result in perceptible watermarks that negatively affect user experience and are easily detectable, making them ineffective against piracy.
Innovation Solution
Employing machine-learned models for imperceptible spatial-temporal watermarking by training a watermark encoding model to embed messages in video frames using a discriminator model to reduce perceptibility and a decoding model to reconstruct the message, while utilizing a video distortion model to enhance robustness against real-world distortions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional watermarking methods are used to embed messages in video data, then the watermark can be detected and extracted, but the watermark becomes perceptible and easily detectable by users and attackers
Solution Approach 1:
The patent transforms the watermark embedding problem from traditional spatial domain methods to frequency domain representations (DCT, DFT, wavelet transforms). By changing the domain parameter where watermarking occurs, the system achieves imperceptibility while maintaining detectability through machine learning models that learn optimal frequency-domain embedding strategies.
Solution Approach 2:
The patent replaces traditional mechanical/mathematical watermark embedding algorithms with machine learning-based models (autoencoders, GANs, CNNs). These learned models adaptively embed watermarks in ways that are imperceptible to humans yet detectable by the system, substituting rigid algorithmic approaches with flexible neural network-based solutions.
2Difficulty of detecting and measuring
If watermark strength is increased to make watermarks more detectable, then the watermark can be easier to detect, but the watermark becomes more perceptible and degrades video quality
Solution Approach 1:
The patent applies different watermark embedding strengths and strategies to different regions and frequency components of the video. Machine learning models identify optimal locations and frequencies for watermark embedding, applying stronger watermarks in less perceptible areas and weaker watermarks in sensitive areas, achieving both detectability and quality preservation.
Solution Approach 2:
The patent uses dynamic, adaptive watermark embedding where the strength and location of watermarks are adjusted based on local video characteristics. Machine learning models analyze video content and adaptively modulate watermark parameters in real-time, making the watermarking process dynamic rather than static, thereby balancing detectability and quality.
3Manufacturing precision
If simple watermarking methods are used to maintain video quality, then the video remains high quality, but the watermark becomes easily detectable and vulnerable to removal
Solution Approach 1:
The patent moves watermarking from the simple spatial domain to multi-dimensional frequency domains (spatial-frequency, time-frequency). By embedding watermarks across multiple dimensions and transformations, the system achieves robust protection while maintaining visual quality, as attackers would need to manipulate multiple dimensions simultaneously to remove the watermark.
Solution Approach 2:
The patent combines multiple watermarking techniques and machine learning models into a composite system. Rather than using a single simple method, the system integrates various embedding strategies, transformation domains, and neural network architectures to create a robust composite watermarking solution that maintains quality while enhancing protection reliability.
Data Source
Figure 1A
Figure 1B
Figure 1C
AI summary
Systems and methods of the present disclosure are directed to a computing system. The computing system can obtain a message vector and video data comprising a plurality of video frames. The computing system can process the input video with a transformation portion of a machine-learned watermark encoding model to obtain a three-dimensional feature encoding of the input video. The computing system can process the three-dimensional feature encoding of the input video and the message vector with an embedding portion of the machine-learned watermark encoding model to obtain spatial-temporal watermark encoding data descriptive of the message vector. The computing system can generate encoded video data comprising a plurality of encoded video frames, wherein at least one of the plurality of encoded video frames includes the spatial-temporal watermark encoding data.