中文版本

Opportunity 

The rapid growth of high-resolution video content and the increasing demand for efficient video compression have highlighted significant challenges in existing video coding standards. Current rate control algorithms, such as those in VVC (Versatile Video Coding), often rely on encoding statistics from previous frames to infer model parameters for the current frame. This approach assumes that current and previous frames share similar content and reference relationships, which may not hold true for videos with large motion or extended Group-of-Pictures (GOP) sizes. As a result, these algorithms struggle to maintain optimal rate-distortion performance, leading to suboptimal bit-rate allocation and quality fluctuations. The need for a more accurate and adaptive rate control mechanism that can dynamically adjust to varying video content is the primary motivation behind this patent.  

Technology

This patent introduces a neural network-based rate control algorithm designed to enhance video encoding efficiency. The innovation lies in leveraging pre-analysis frameworks and deep learning to predict rate-distortion characteristics more accurately. The method involves:  

1. Pre-analysis Framework: A proxy encoding process extracts prediction residuals from video data units (e.g., Coding Tree Units, CTUs) using limited partition modes (e.g., quad-tree) to balance complexity and accuracy.  
2. Neural Network-Based Modeling: Four neural networks (two for intra-frame and two for inter-frame coding) process these residuals to estimate model parameters (α for bit-rate, β for distortion) for a rate-distortion model. The networks share a common structure with feature extractors and regressors.  
3. Dynamic Refinement: Model parameters are adjusted using refinement factors based on actual encoding outcomes, ensuring adaptability to content variations.  
4. Optimal Bit Allocation: Frame- and CTU-level bit allocation is optimized using derived parameters, followed by quantization step size and coding parameter determination (e.g., QP, λ).  

The technology integrates seamlessly with VVC standards (e.g., VTM-13.0), offering improved accuracy over traditional hyperbolic models.

Advantages 

• Higher Accuracy: Neural networks predict rate-distortion parameters directly from residuals, reducing reliance on historical data assumptions.  
• Adaptability: Refinement factors dynamically adjust model parameters based on real-time encoding feedback.  
• Performance Gains: Achieves up to 1.77% BD-Rate savings under Random Access configurations compared to default VTM-13.0 algorithms.  
• Stability: More stable bit-rate output and buffer status, especially in dynamic scenes or large GOPs.  
• Low Latency: Pre-analysis framework minimizes additional computational overhead by limiting partition modes and disabling non-essential coding tools during proxy encoding. 

Applications 

• Video Compression Standards: Integration with VVC/H.266, HEVC/H.265, or future codecs for enhanced rate control.  
• Streaming Services: Improved quality-bitrate trade-offs for platforms like Netflix, YouTube, or teleconferencing tools (Zoom, Teams).  
• Broadcast & Storage: Efficient encoding for UHD/4K broadcasts or archival systems with constrained bandwidth/storage.  
• Edge Devices: Real-time optimization for mobile devices or IoT cameras with limited resources.