Regularized Convolutional Block Prediction for Chroma From Luma
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video encoding and decoding methods struggle to effectively leverage spatial and temporal redundancy in video content for efficient compression, particularly in predicting chroma components from luma components, leading to suboptimal performance in video coding schemes.
Innovation Solution
The method involves deriving filter coefficients using a regularized autocorrelation matrix and cross-correlation vector to generate predictions for chroma samples based on luma samples, employing techniques like convolutional cross-component models and extrapolation filters, with regularization to stabilize the prediction process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional prediction methods are used to predict chroma components from luma components, then the encoding process is simple, but the compression efficiency is suboptimal
Solution Approach 1:
The patent applies parameter changes by introducing regularization parameters (lambda values) to modify the prediction model. The regularization term is added to the optimization objective function, transforming the conventional prediction approach into a regularized one that balances prediction accuracy and model stability. This allows the system to achieve better compression efficiency by optimizing the prediction parameters without excessively increasing complexity.
Solution Approach 2:
The patent introduces an intermediary regularization mechanism that mediates between the prediction accuracy and model stability. The regularization term acts as an intermediary component that guides the filter coefficient optimization process, enabling the system to leverage spatial and temporal redundancy more effectively while maintaining computational tractability.
2Measurement precision
If regularization is applied to stabilize the prediction process, then the prediction accuracy improves, but the computational complexity increases
Solution Approach 1:
The patent applies local quality by allowing different regularization parameters for different regions or components. The regularization matrix can have different diagonal elements, enabling localized control over the prediction stability. This approach improves prediction accuracy in critical areas while maintaining computational efficiency by applying stronger regularization only where needed.
Solution Approach 2:
The patent implements partial regularization by selectively applying regularization to specific filter coefficients or prediction components. Not all parameters are regularized with the same strength, allowing the system to achieve sufficient prediction accuracy while minimizing the computational overhead associated with full regularization.
3Manufacturing precision
If filter coefficients are derived using regularized autocorrelation matrix, then the prediction of chroma samples improves, but the computational load increases
Solution Approach 1:
The patent applies segmentation by dividing the filter coefficient derivation process into manageable stages. The autocorrelation matrix computation and regularization are performed separately, and the filter coefficients are derived in a structured manner. This segmentation allows for optimized computation at each stage, improving prediction precision while controlling the overall computational load.
Solution Approach 2:
The patent uses copying by reusing the regularized autocorrelation matrix across different prediction scenarios. Once the matrix is computed and regularized, it can be copied and applied to multiple chroma sample predictions, reducing redundant computations and lowering the overall computational load while maintaining high prediction precision.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
In one implementation, the encoder or decoder first obtains a first type and a second type of samples in a template area of a current block. From the first type of samples, the encoder or decoder obtains a matrix indicating autocorrelation of the first type of samples in the template area. The matrix is adjusted by a regularization matrix to form a regularized autocorrelation matrix, wherein non-zero elements of a diagonal of the regularization matrix are allowed to be different. The encoder or decoder obtains a vector indicating cross-correlation of the first type of samples and the second type of samples in the template area. Based on the regularized autocorrelation matrix and the cross-correlation vector, the encoder or decoder obtains a set of filter coefficients to be used to generate a prediction for a second type of samples in the current block based on a first type of samples.