Line Graph Transform Coding With Reduced 32-Point and 64-Point Complexity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video coding technologies, such as AV1 and AV2, face high computational complexity due to the use of large size transforms like 32-point and 64-point Line Graph Transforms (LGTs), particularly in direct matrix multiplications, which are resource-intensive.
Innovation Solution
Implement methods to reduce computational complexity by zeroing out high-frequency coefficients and employing non-separable secondary transforms, such as non-separable secondary transforms (NSST) and optimized Line Graph Transforms (LGTs) for video coding, specifically for 32-point and 64-point transforms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large size transforms (32-point and 64-point LGTs) are used for video coding, then transform accuracy and compression performance are improved, but computational complexity increases significantly
Solution Approach 1:
The transform block is segmented into multiple sub-blocks, and the transform process is divided into primary transform and secondary transform stages. This segmentation allows the large 32-point and 64-point LGTs to be processed in smaller, more manageable units, reducing the computational burden while maintaining transform accuracy through multi-stage processing.
Solution Approach 2:
Different transform types are applied to different regions based on local characteristics. The patent applies non-separable secondary transforms (NSST) specifically to directional texture patterns, while using conventional transforms for other regions. This local adaptation optimizes compression performance for specific image features without requiring full complex processing everywhere.
2Measurement precision
If direct matrix multiplications are used for 32-point and 64-point LGTs, then transform precision is maintained, but resource consumption increases
Solution Approach 1:
The patent modifies the transform parameters by introducing a secondary transform stage that operates on the results of the primary transform. By changing the processing parameters from a single large matrix multiplication to a two-stage process with different matrix sizes, the computational resources required are reduced while maintaining or improving transform precision through the combined effect of both stages.
3Speed
If conventional transforms are used for all blocks, then processing speed is maintained, but coding efficiency decreases for directional texture patterns
Solution Approach 1:
The transform type is made dynamic and adaptive rather than fixed. The patent introduces a secondary transform that is conditionally applied based on the characteristics of the residual block, particularly for directional texture patterns. This dynamic approach allows the system to switch between conventional and enhanced transform methods, improving coding efficiency for specific patterns while maintaining overall processing speed through selective application.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method of coding image data performed by at least one processor, may include: receiving information regarding a data block of an image; determining whether at least one of a height or a width of a residual coding block corresponding to the data block of the image is greater than or equal to a pre-defined threshold; and based on determining that the at least one of the height or the width of the residual coding block is greater than or equal to the pre-defined threshold: identifying or generating a reduced-complexity residual coding block by reducing the number of non-zero coefficients in the residual coding block; and performing transform coding of the reduced-complexity residual coding block using a line graph transform (LGT) core to perform direct matrix multiplications for each of the horizontal and vertical dimensions of the reduced-complexity coding block.