Decoding Curved Basis Transforms Flat Residual Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image and video decoding technologies using transforms with non-flat lowest-frequency basis images, such as KLT and DST, often generate artifacts in smooth or slowly changing areas due to their curved or concave shapes, leading to degradation in subjective quality and compression efficiency.
Innovation Solution
A decoder method that reconstructs residual blocks with only the lowest-frequency coefficient as a flat image, depending on its value, and applies an inverse spatial 2D transform for blocks with multiple coefficients, thereby avoiding the use of non-flat basis images that cause artifacts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If transforms with non-flat lowest-frequency basis images (KLT, DST) are used, then compression efficiency is improved, but artifacts are generated in smooth or slowly changing areas
Solution Approach 1:
The patent applies different reconstruction strategies based on the characteristics of the residual block: flat reconstruction is used for blocks with only DC coefficient (affecting smooth areas), while inverse transform is used for blocks with multiple coefficients (affecting detailed areas). This local differentiation resolves the contradiction by applying the appropriate method to each region's needs.
Solution Approach 2:
Instead of using the conventional inverse transform for all blocks, the patent inverts the approach by using a simplified flat reconstruction method specifically for blocks with only DC coefficient, thereby eliminating artifacts in smooth areas while maintaining compression efficiency.
2Productivity
If KLT is used to optimize energy compaction, then compression efficiency is improved, but computational complexity and parameter communication requirements increase
Solution Approach 1:
The patent replaces the complex KLT basis functions with a simple flat reconstruction approach for DC-only blocks, using a computationally inexpensive method that achieves the same goal of representing smooth areas without requiring complex calculations or parameter communication.
Solution Approach 2:
The patent extracts and handles the DC coefficient case separately from the general transform case, removing the need to apply complex KLT basis functions when only the DC component is present, thereby reducing computational complexity while maintaining compression efficiency.
3Productivity
If DST is used for signals with negative correlation, then compression efficiency is improved, but curved basis images generate artifacts in flat areas
Solution Approach 1:
The patent applies flat reconstruction specifically to blocks with only DC coefficient that correspond to smooth or flat areas, while allowing DST to be used for blocks with multiple coefficients that represent detailed areas with negative correlation. This local differentiation eliminates artifacts in flat areas while preserving DST's compression benefits for appropriate content.
Solution Approach 2:
The patent introduces dynamic selection of reconstruction methods based on the number of non-zero coefficients in the transform block, adapting the reconstruction strategy to the local characteristics of each block rather than applying a fixed method universally.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
The solution according to embodiments of the present invention is to reconstruct image blocks where only the frequency coefficient that is most sensitive to changes of an average value of the residual image block, typically the lowest frequency coefficient, is present with a flat residual image. The other image blocks are reconstructed by applying an inverse spatial 2D transform. Hence, the other image blocks comprise image blocks with multiple coefficients and image blocks with a single coefficient and wherein the single coefficient is not the coefficient determining (703) that is most sensitive to changes of an average value of the residual image block.