DIMD Position-Dependent Blending for Accurate Intra Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional decoder-side intra mode derivation (DIMD) processes in video coding fail to account for local characteristics of intra prediction modes, leading to less accurate predictors and additional computational overhead due to global blending weights.
Innovation Solution
A method and apparatus for deriving a DIMD predictor that involves selecting intra prediction modes (IPMs) in a template area adjacent to the coding unit, determining blending weights based on the distance of samples to the template area, and generating the predictor by blending the selected IPMs with Planar or DC mode using these weights.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If global blending weights are used in DIMD process, then the implementation is simpler, but the prediction accuracy deteriorates due to ignoring local characteristics
Solution Approach 1:
The patent applies local quality by dividing the template area into multiple regions and assigning different blending weights to each region based on its local characteristics. Specifically, the template area is divided into a first region and a second region, with first blending weights applied to samples in the first region and second blending weights applied to samples in the second region. This allows the prediction to adapt to local variations in the template area, improving prediction accuracy while maintaining a relatively simple implementation structure.
2Measurement precision
If position-dependent blending weights are implemented, then prediction accuracy improves, but computational overhead increases
Solution Approach 1:
The patent reduces computational overhead through segmentation by dividing the template area into a limited number of regions (first region and second region). This segmentation approach allows the use of position-dependent blending weights without requiring complex calculations for each individual sample. The blending weight determination module can efficiently select appropriate weights based on the sample's region, significantly reducing the computational complexity compared to calculating unique weights for every sample position.
Solution Approach 2:
The patent employs parameter changes by using a limited set of blending weight parameters corresponding to different regions. Instead of calculating continuous or highly variable weights, the system changes the weight parameter discretely based on the sample's region classification. This approach maintains high prediction accuracy while keeping the computational overhead manageable through parameter optimization and reduction.
3Productivity
If uniform blending is applied across the coding unit, then the processing is faster, but the fidelity of intra prediction deteriorates
Solution Approach 1:
The patent resolves this contradiction by applying local quality through region-specific blending weights. The template area is divided into distinct regions, each with its own blending weight characteristics. Samples in different regions are blended using weights that reflect their local properties, thereby improving intra prediction fidelity while maintaining processing efficiency through the structured regional approach.
Solution Approach 2:
The patent applies partial action by implementing position-dependent blending weights only for samples that benefit from it, rather than uniformly across all samples. The blending weight determination module selectively applies different weighting strategies based on the sample's position and region, achieving high fidelity where needed while maintaining fast processing for other samples, thus balancing productivity and prediction fidelity.
Data Source
AI summary
A method of deriving a Decoder-side Intra Mode Derivation (DIMD) predictor for respective samples of a coding unit (CU) of a picture, includes: selecting one or more Intra Prediction Modes (IPMs) in a template area adjacent to the CU; for each sample of the CU, determining, dependent on a distance of the sample of the CU to the template area, blending weights for blending at least the one or more selected IPMs; and generating the DIMD predictor by blending the one or more selected IPMs and a Planar or DC mode using the determined blending weights.