Meta-Learned Neural Image Compression for Smooth Bitrate Control

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing neural image compression methods struggle with flexible bitrate control, requiring multiple model instances for different trade-offs between rate and distortion, and lack the ability to achieve arbitrary smooth bitrate control due to the difficulty in training and storing an infinite number of model instances for every possible target bitrate.

Innovation Solution

A meta-learning mechanism is employed to adaptively compute rate-adaptive weight parameters of a neural image compression model, allowing a single model instance to achieve image compression with arbitrary smooth bitrates by separating model parameters into shared and adaptive components, and using a meta-learning framework to train these parameters for various bitrate requirements.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If multiple model instances are trained for different bitrate trade-offs, then rate-distortion control capability is improved, but device complexity and storage requirements increase

Engineering Contradiction:
Improverate-distortion control capabilityVSAvoidnumber of model instances
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent implements a universal neural network model that can perform multiple rate-distortion optimization tasks across different bitrates. The model uses shared encoding and decoding neural networks that are trained to handle various bitrate requirements through a unified training process, eliminating the need for separate model instances for each bitrate target.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The patent changes the approach from training multiple models with fixed parameters to training a single model with learnable parameters that can adapt to different bitrates. The model uses bitrate-dependent weight parameters that are optimized during training to achieve rate-distortion optimization across a range of bitrates, allowing flexible bitrate control without increasing model quantity.

Inventive Principle:
Principle #35Parameter changes

2Device complexity

If a single model instance is used for multiple bitrates, then device complexity is reduced, but arbitrary smooth bitrate control capability deteriorates

Engineering Contradiction:
Improvenumber of model instancesVSAvoidarbitrary smooth bitrate control
Core Design Contradiction:
Device complexityVSAdaptability or versatility

Solution Approach 1:

The patent introduces dynamic adaptability into the single model instance by implementing bitrate-dependent weight parameters that can be adjusted based on the target bitrate. The model uses learnable weight parameters that are optimized during training to achieve smooth bitrate control, allowing the system to adapt to arbitrary bitrate requirements without requiring multiple fixed models.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent performs preliminary training of the model with bitrate-dependent weight parameters during the training phase. The model is pre-trained to learn optimal weight configurations for different bitrate ranges, enabling it to achieve smooth bitrate control during inference without requiring retraining or multiple model instances for different bitrate targets.

Inventive Principle:
Principle #10Preliminary action

3Adaptability or versatility

If multiple model instances are deployed, then coverage of bitrate demands is improved, but storage and deployment requirements increase

Engineering Contradiction:
Improvecoverage of bitrate demandsVSAvoidstorage of model instances
Core Design Contradiction:
Adaptability or versatilityVSQuantity of substance

Solution Approach 1:

The patent deploys a single universal model instance that covers a wide range of bitrate demands through its ability to adapt to different bitrates during inference. The model uses shared encoding and decoding neural networks with bitrate-dependent weight parameters that allow it to serve multiple bitrate requirements, significantly reducing storage requirements compared to deploying multiple separate model instances.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The patent merges the functionality of multiple bitrate-specific models into a single unified model instance. By combining shared encoding and decoding neural networks with bitrate-dependent weight parameters, the system achieves comprehensive bitrate coverage while storing and deploying only one model, reducing the quantity of model data required.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP4115610B1Method and apparatus for adaptive neural image compression with rate control by meta-learning
Publication Date: 2025.10.15 TENCENT AMERICA LLC
  • EP4115610B1 patent drawingFigure 1
  • EP4115610B1 patent drawingFigure 2
  • EP4115610B1 patent drawingFigure 3A

AI summary

A method of adaptive neural image compression with rate control by meta-learning includes receiving an input image and a hyperparameter; and encoding the received input image, based on the received hyperparameter, using an encoding neural network, to generate a compressed representation. The encoding includes performing a first shared encoding on the received input image, using a first shared encoding layer having first shared encoding parameters, performing a first adaptive encoding on the received input image, using a first adaptive encoding layer having first adaptive encoding parameters, combining the first shared encoded input image and the first adaptive encoded input image, to generate a first combined output, and performing a second shared encoding on the first combined output, using a second shared encoding layer having second shared encoding parameters.