Partitioned INR Encoding for Efficient Neural Signal Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural compression techniques, particularly INR-based methods, face challenges in efficiently exploiting spatial and temporal redundancies in neural network parameters, leading to high computational complexity and suboptimal coding efficiency.
Innovation Solution
The proposed method involves partitioning the signal into multiple INR partitions, converting these parameters into parameter images or videos using traditional or machine learning-based codecs like VVC or HEVC, and encoding them to exploit spatial and temporal redundancies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If INR-based compression techniques are used, then computational complexity is reduced, but coding efficiency deteriorates due to inability to exploit spatial and temporal redundancies
Solution Approach 1:
The patent divides the signal into multiple partitions and represents each partition using separate INR parameters. This segmentation allows traditional video coding tools to be applied to each partition independently, enabling exploitation of spatial and temporal redundancies while maintaining the low computational complexity benefit of INR-based methods.
Solution Approach 2:
The patent merges the advantages of two different approaches: it combines the low computational complexity of INR-based methods with the high coding efficiency of traditional video compression standards like VVC or HEVC. By integrating INR parameter representation with traditional coding tools, the patent achieves both reduced computational complexity and improved coding efficiency.
2Productivity
If traditional neural compression techniques are used, then coding efficiency is maintained, but computational complexity increases
Solution Approach 1:
The patent segments the signal into partitions and uses INR parameters for each partition, which can be efficiently encoded using traditional video coding techniques. This segmentation enables the use of optimized coding tools while keeping the computational complexity manageable through the structured partitioning approach.
Solution Approach 2:
The patent introduces an intermediary representation layer where INR parameters serve as a bridge between the original signal and the compressed bitstream. This intermediary form allows traditional video coding tools to operate effectively on the parameters rather than the raw signal, improving coding efficiency while maintaining computational tractability.
3Device complexity
If INR parameters are directly encoded without partitioning, then device complexity is reduced, but information loss increases due to inability to exploit redundancies
Solution Approach 1:
The patent segments the INR parameters according to signal partitions and applies traditional video coding tools to each segment. This segmentation preserves important spatial and temporal information by enabling redundancy exploitation, thereby reducing information loss while maintaining relatively low device complexity through the use of standardized coding tools.
Data Source
Figure 1
Figure 2~3
Figure 4~5
AI summary
In one implementation, an image including multiple sets of INR (Implicit Neural Representation) parameters is decoded, wherein at least a sample in the decoded image represents an INR parameter; multiple partitions are decoded for the signal, wherein each set of the multiple sets of INR parameters corresponds to a respective partition of the multiple partitions of the signal; and the signal is reconstructed based on the multiple partitions of the signal. On the corresponding encoder side, the signal is partitioned into multiple partitions; a set of INR parameters is obtained to represent each partition of the multiple partitions of the signal; each set of INR parameters is converted to a corresponding region in an image, wherein at least a sample in the corresponding region represents an INR parameter of the set of INR parameters; and the image is encoded.