Matrix-Based Intra-Prediction Using Multi-Reference Lines
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing intra-prediction technologies, such as Matrix Based Intra Prediction (MIP), suffer from low prediction accuracy due to the limited use of reference pixel information, resulting in significant image information loss.
Innovation Solution
The method involves determining multiple reference lines for a current coding block, down-sampling these lines to obtain second reference pixels, performing matrix multiplication to generate a reference matrix, and using this matrix along with first reference pixels to create a matrix-based intra-prediction block, which improves prediction accuracy by incorporating more image information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional MIP mode is used with single reference line, then device complexity is reduced, but prediction accuracy deteriorates due to information loss
Solution Approach 1:
The current coding block is divided into multiple regions (first prediction block and second prediction block) that use different prediction methods. The first prediction block uses traditional MIP mode while the second prediction block uses the enhanced multi-reference line MIP mode, allowing selective application of complex processing only where beneficial for prediction accuracy.
Solution Approach 2:
Instead of applying the enhanced multi-reference line processing to all coding blocks, the method selectively applies it only to the second prediction block when the block type indicates enhanced prediction is needed. This partial application reduces overall complexity while maintaining accuracy improvements where most beneficial.
2Loss of information
If multiple reference lines are used with down-sampling, then prediction accuracy is improved by retaining more image information, but computational complexity increases
Solution Approach 1:
Reference lines are down-sampled before matrix multiplication to pre-process and compress the reference pixel data. This preliminary down-sampling reduces the amount of data that needs to be processed in subsequent matrix operations, retaining essential image information while reducing computational complexity.
Solution Approach 2:
The reference lines undergo parameter transformation through down-sampling, changing their resolution and data density. This parameter change allows multiple reference lines to be processed efficiently by reducing the number of pixels while preserving the essential image information needed for accurate prediction.
3Measurement precision
If down-sampling of reference lines is performed, then more reference pixel information is utilized, but processing time increases
Solution Approach 1:
Down-sampling and enhanced processing is applied only to the second prediction block when block type indicates enhanced prediction is needed, rather than processing all blocks with the same complexity. This selective approach reduces overall processing time while maintaining accuracy improvements where most beneficial.
Data Source
AI summary
A method for intra-prediction is provided, including: determining at least one reference line of a current coding block, each of the at least one reference line comprising a plurality of first reference pixels; down-sampling the at least one reference line and obtaining a plurality of second reference pixels; performing matrix multiplication for the plurality of second reference pixels and obtaining a reference matrix; and obtaining a matrix-based intra-prediction block of the current coding block based on the reference matrix and the plurality of first reference pixels, the matrix-based intra-prediction block comprising a plurality of first pixels and a plurality of second pixels, a value of each of the first pixels is a value of a corresponding pixel in the reference matrix, and a value of each of the second pixels is obtained based on values of the first pixels and values of the first reference pixels.