Video Decoder MIP LFNST Kernel Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The use of decoder side intra mode derivation (DIMD) technology in video coding introduces complexity in both software and hardware, increasing compression costs and reducing encoding efficiency.
Innovation Solution
A method that decodes a bitstream to determine prediction mode parameters, including matrix-based intra prediction (MIP) and low-frequency non-separable transform (LFNST) parameters, to select appropriate transform kernels and reduce computational complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If decoder side intra mode derivation (DIMD) technology is used to derive intra prediction mode at decoding side, then bit overhead is saved, but software and hardware complexity increases significantly
Solution Approach 1:
The patent changes the parameter of block size classification, dividing blocks into small blocks (4×4, 8×8) and large blocks (16×16, 32×32). For small blocks, DIMD technology is applied to derive intra prediction modes and save bit overhead. For large blocks, traditional methods are used to avoid excessive complexity. This parameter-based segmentation resolves the contradiction by applying the advanced technique only where it provides net benefit.
Solution Approach 2:
The patent applies different processing qualities to different regions (block sizes). Small blocks receive the sophisticated DIMD treatment with full derivation capabilities, while large blocks receive simplified traditional prediction. This local differentiation ensures that complexity is concentrated where it most effectively reduces bit overhead, without unnecessarily complicating the processing of larger blocks.
2Productivity
If DIMD technology is applied to all block sizes, then compression efficiency improves, but encoding cost increases
Solution Approach 1:
The patent modifies the application scope parameter of DIMD technology based on block size. By setting a threshold (small blocks only), it achieves optimal compression efficiency for the blocks that benefit most while capping the encoding cost. This selective application resolves the contradiction between improving compression efficiency and controlling encoding cost.
Solution Approach 2:
Instead of applying DIMD to all blocks (excessive action), the patent applies it partially only to small blocks where the benefit-to-cost ratio is most favorable. This partial application achieves sufficient compression efficiency improvement without incurring excessive encoding costs across all block types.
3Measurement precision
If LFNST transform kernel candidate sets are expanded to improve transform accuracy, then prediction precision improves, but computational complexity increases
Solution Approach 1:
The patent changes the parameter of LFNST kernel selection based on MIP mode parameters and block characteristics. Instead of using a fixed large set of kernel candidates for all blocks, the selection is adapted dynamically based on the specific MIP mode and block properties. This resolves the contradiction by providing high precision where needed while reducing computational complexity through selective candidate set usage.
Solution Approach 2:
Different LFNST kernel candidate sets are applied to different local conditions (different MIP modes and block types). This local adaptation ensures that prediction precision is maximized for each specific case without requiring the system to maintain and process all possible kernel candidates for every block, thereby controlling overall computational complexity.
Data Source
AI summary
A decoding method is provided. A bitstream is decoded to determine a prediction mode parameter. The bitstream is decoded to determine a matrix-based intra prediction (MIP) parameter of a current block, in response to the prediction mode parameter indicating that MIP is used to determine an intra prediction value. The bitstream is decoded to determine transform coefficients of the current block and a low-frequency non-separable transform (LFNST) index of the current block. A mapping mode of an LFNST transform set is determined according to the MIP parameter, in response to the LFNST index indicating that an LFNST is used for the current block. According to the mapping mode of the LFNST transform set, one LFNST transform kernel candidate set is selected from a plurality of LFNST transform kernel candidate sets, and an LFNST transform kernel used for the current block is determined from the selected LFNST transform kernel candidate set.