Image Coding With Block-Size Adaptive LFNST Kernels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The increasing demand for high-resolution and high-quality images/videos, particularly in immersive media formats like VR and AR, necessitates a highly efficient image/video compression technique to minimize transmission and storage costs while maintaining coding efficiency.
Innovation Solution
An image coding method and apparatus that applies LFNST (Low-Frequency Non-Separable Transform) to adjust the zero-out area, considering computational complexity, and sets the LFNST kernel based on the target block's dimensions, deriving modified transform coefficients through specific matrix operations to enhance coding performance and minimize complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If LFNST is applied with fixed transform kernel sizes, then the coding process is simple, but coding efficiency is limited for various block sizes
Solution Approach 1:
The patent implements dynamic selection of LFNST transform kernel sizes based on the actual block dimensions. The transform kernel size is adjusted according to whether the block width and height are greater than or equal to 8 pixels, enabling adaptive optimization for different block sizes while maintaining systematic control over the transformation process
Solution Approach 2:
The patent changes the transform kernel parameters (size and type) based on block characteristics. Specifically, it selects between different kernel sizes (e.g., 4×4, 8×8, 16×16) and transform types (DCT-II, DST-VII) according to the block width and height, optimizing the transformation parameters to match the input data characteristics
2Productivity
If larger transform kernels are used for larger blocks, then coding performance improves, but computational complexity increases
Solution Approach 1:
The patent adapts the transform kernel size parameter to match the block size, using larger kernels (e.g., 16×16) for larger blocks and smaller kernels (e.g., 4×4, 8×8) for smaller blocks. This parameter adaptation optimizes compression efficiency while avoiding the computational overhead of using excessively large kernels for small blocks
Solution Approach 2:
The patent applies transform kernels with sizes that are appropriate but not excessive for each block. By selecting kernel sizes that match the block dimensions rather than always using the largest available kernel, it achieves sufficient compression performance while minimizing unnecessary computational complexity
3Productivity
If LFNST is applied to all blocks, then coding efficiency is maximized, but the amount of signaling data increases
Solution Approach 1:
The patent applies LFNST selectively to specific blocks based on their characteristics (size, transform type, prediction mode) rather than uniformly to all blocks. This localized application approach maintains coding efficiency for blocks that benefit from LFNST while avoiding unnecessary processing and signaling overhead for blocks where LFNST provides minimal benefit
Data Source
AI summary
An image decoding method according to this document comprises a step of deriving, for transform coefficients, modified transform coefficients on the basis of inverse secondary transform, wherein: the step of deriving the modified transform coefficients may comprise a step of deriving a transform kernel to be applied to the inverse secondary transform; on the basis of both the horizontal and vertical lengths of a target block being greater than or equal to 8 and the horizontal or vertical length being 8, the transform kernel may be set to a 64×32 matrix; and on the basis of both the horizontal and vertical lengths of the target block being 8, a 64×16 matrix sampled from the 64×32 matrix may be applied to the inverse secondary transform of the target block.