INR Parameter Tensor Encoding for Lower Neural Network Bitrate

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for encoding Implicit Neural Representation (INR) parameters in bitstreams are sub-optimal when multiple small neural networks are used, leading to inefficient bitrate usage, especially when applying the Neural Network Coding (NNC) standard, which fails to leverage sequential redundancies effectively.

Innovation Solution

The proposed solution involves concatenating the parameters of multiple small INR networks into a single tensor and utilizing context-adaptive binary arithmetic coding (CABAC) to create sequential redundancy, optimizing the encoding process.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If multiple small INR networks are encoded separately using traditional methods, then each network can be processed independently, but the bitrate efficiency deteriorates significantly

Engineering Contradiction:
Improveencoding processabilityVSAvoidbitrate efficiency
Core Design Contradiction:
Ease of manufactureVSLoss of energy

Solution Approach 1:

The patent merges multiple small INR network parameters into a single concatenated tensor structure, allowing them to be encoded together as one unified data structure. This combining approach enables the encoder to leverage sequential redundancies across all networks simultaneously, achieving up to 80% bitrate reduction compared to independent encoding while maintaining the ability to process each network through the unified structure.

Inventive Principle:
Principle #5Merging (Combining)

2Ease of operation

If the Neural Network Coding standard is applied directly to multiple small INR networks, then the encoding process is simple, but sequential redundancies are not leveraged effectively

Engineering Contradiction:
Improveencoding simplicityVSAvoidredundancy utilization
Core Design Contradiction:
Ease of operationVSLoss of information

Solution Approach 1:

The patent performs preliminary action by reordering and concatenating the parameters of multiple small INR networks into a specific sequence before encoding. This preliminary arrangement creates sequential redundancies that the CABAC encoder can then effectively exploit. The reordering strategy prepares the data in advance to maximize redundancy utilization, achieving better compression without complicating the actual encoding operation.

Inventive Principle:
Principle #10Preliminary action

3Loss of energy

If parameters of multiple INR networks are concatenated into a single tensor, then sequential redundancies are created for better compression, but the data structure becomes more complex

Engineering Contradiction:
Improvebitrate efficiencyVSAvoiddata structure complexity
Core Design Contradiction:
Loss of energyVSDevice complexity

Solution Approach 1:

The patent applies segmentation by dividing the data into multiple small INR network parameters that are then systematically concatenated into a single tensor. The segmentation allows each network to maintain its structural integrity while enabling the concatenated structure to benefit from sequential redundancies. This segmented-then-combined approach achieves bitrate efficiency without requiring complete redesign of the data structure, as the tensor maintains standard multi-dimensional array organization.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4664881A1Efficient compression of coding tree unit based implicit neural representation with neural network coding standard
Publication Date: 2025.12.17 INTERDIGITAL CE PATENT HOLDINGS SAS
  • EP4664881A1 patent drawingFigure 1A~1B
  • EP4664881A1 patent drawingFigure 1C
  • EP4664881A1 patent drawingFigure 2~4

AI summary

A method and an apparatus for encoding or decoding an image, a video or a 3D object are provided wherein data representative of the image, video or 3D object is partitioned into coding units. For each coding units, parameters of Implicit Neural Representation network are obtained, wherein each one of the coding units is represented using one of the Implicit Neural Representation networks. The obtained parameters are concatenated into into at least one tensor and encoded using context-adaptive binary arithmetic coding.