LFNST Kernel Grouping for Video Coding Memory Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current video coding techniques face inefficiencies in memory storage due to the need for separate low frequency non-separable transform (LFNST) kernels for various transform unit sizes and intra prediction modes, leading to increased memory overhead.
Innovation Solution
The proposed techniques group transform units into classes based on size and intra prediction modes, allowing a single LFNST kernel to be applied across multiple units and modes, reducing the number of kernels required and minimizing memory storage needs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If separate LFNST kernels are stored for each transform unit size and intra prediction mode, then the transform can be optimized for each specific case, but memory overhead increases significantly
Solution Approach 1:
The patent merges multiple LFNST kernels into a single unified kernel by applying context-dependent scaling factors. Instead of storing separate kernels for different transform unit sizes and prediction modes, the system stores one base kernel and multiplies it by scaling factors selected based on the specific context (TU size and prediction mode), thereby reducing memory storage requirements while maintaining optimization benefits.
Solution Approach 2:
The unified LFNST kernel is designed to serve multiple functions across different transform unit sizes and intra prediction modes. By making the kernel universal and using scaling factors to adapt it to specific contexts, the system eliminates the need for multiple specialized kernels, thus reducing memory overhead while preserving the ability to optimize transforms for various cases.
2Manufacturing precision
If multiple LFNST kernels are maintained for different transform unit sizes, then each size can use an optimized kernel, but the complexity of kernel selection and management increases
Solution Approach 1:
The patent combines multiple kernels into one unified kernel structure. Instead of managing separate kernels for different transform unit sizes, the system uses a single kernel with context-dependent scaling factors, thereby reducing the complexity of kernel selection and management while maintaining transform precision through the scaling mechanism.
Solution Approach 2:
The system changes the approach from selecting different kernels to changing scaling factors applied to a single kernel. This parameter change simplifies kernel management because instead of managing multiple kernel matrices, the system only needs to manage a set of scaling factors that are selected based on transform unit size and prediction mode.
3Reliability
If separate LFNST kernels are stored for each intra prediction mode, then mode-specific optimization is achieved, but memory storage requirements increase
Solution Approach 1:
The patent merges LFNST kernels across different intra prediction modes by using a unified kernel with mode-specific scaling factors. Instead of storing separate kernels for each prediction mode, the system stores one kernel and selects appropriate scaling factors based on the prediction mode, thereby reducing memory storage requirements while maintaining mode-specific optimization.
Solution Approach 2:
The unified LFNST kernel is made universal to work across all intra prediction modes. By combining the kernel with mode-dependent scaling factors, the system enables a single kernel to serve multiple prediction modes effectively, reducing the need for separate kernels and thus reducing memory storage requirements.
Data Source
AI summary
A method of decoding video data includes determining an intra prediction mode from a plurality of intra prediction modes for a current block of the video data, determining a low frequency non-separable transform (LFNST) kernel from a plurality of LFNST kernels for the current block based on the determined intra prediction mode, wherein at least one LFNST kernel of the plurality of LFNST kernels is associated with at least two different intra prediction modes of the plurality of intra prediction modes, applying an inverse of the determined LFNST kernel to coefficient values generated from a transform unit (TU) of the current block to generate intermediate values, applying an inverse primary transform on the intermediate values to generate residual data, and reconstructing the current block based on the residual data.