Adaptive Neural Image Compression via Meta-Learning

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing neural image compression methods face challenges in achieving flexible bitrate control and quality metric control, requiring multiple model instances for different target bitrates and quality metrics, which is impractical due to the need for infinite model instances for every possible setting and difficulty in smooth quality metric control.

Innovation Solution

The method employs meta-learning to generate a substitute input image and target quality control parameter, using an encoding neural network to compress the image and a decoding neural network to reconstruct it, allowing for adaptive compression with arbitrary smooth bitrate and quality metric control through a single Meta-NIC model instance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If multiple model instances are trained for different target bitrates and quality metrics, then flexible bitrate control and quality metric control are achieved, but device complexity and storage requirements increase significantly

Engineering Contradiction:

Improveflexible bitrate controlVSAvoidnumber of model instances

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent applies universality by training a single neural image compression model to handle multiple target bitrates and quality metrics simultaneously. The model uses learnable parameters that can be dynamically adjusted during inference to achieve different compression rates and quality levels, eliminating the need for multiple specialized model instances while maintaining flexible bitrate and quality control capabilities

Inventive Principle:

Principle #6Universality (Multi-functionality)

Solution Approach 2:

The patent employs parameter changes by introducing learnable parameters within the neural network that can be dynamically modified during inference to achieve different target bitrates and quality metrics. These parameters allow the single model to adapt its compression behavior by adjusting weights and activation thresholds, enabling flexible control without requiring multiple fixed model instances

Inventive Principle:

Principle #35Parameter changes

2Adaptability or versatility

If multiple model instances are trained for different target bitrates and quality metrics, then flexible bitrate control and quality metric control are achieved, but storage requirements increase significantly

Engineering Contradiction:

Improveflexible quality metric controlVSAvoidstorage space

Core Design Contradiction:

Adaptability or versatilityVSQuantity of substance

Solution Approach 1:

Inventive Principle:

Principle #6Universality (Multi-functionality)

Solution Approach 2:

Inventive Principle:

Principle #35Parameter changes

3Device complexity

If a single model instance is used for compression, then device complexity and storage requirements are reduced, but arbitrary smooth bitrate control and quality metric control become difficult

Engineering Contradiction:

Improvenumber of model instancesVSAvoidarbitrary smooth bitrate control

Core Design Contradiction:

Device complexityVSAdaptability or versatility

Solution Approach 1:

The patent applies dynamics by making the model parameters adjustable and adaptive during inference rather than fixed. The neural network incorporates learnable parameters that can be dynamically modified based on the desired target bitrate and quality metric, enabling smooth and arbitrary control over compression characteristics while using a single model instance

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

Inventive Principle:

Principle #35Parameter changes

4Device complexity

If a single model instance is used for compression, then device complexity and storage requirements are reduced, but quality-adaptive weight parameter computation becomes insufficient

Engineering Contradiction:

Improvenumber of model instancesVSAvoidquality-adaptive weight parameter computation

Core Design Contradiction:

Device complexityVSManufacturing precision

Solution Approach 1:

The patent applies feedback by implementing a mechanism that uses the actual compression results to adjust the quality-adaptive weight parameters. The system computes these parameters based on feedback from the compression process, allowing the single model instance to refine its performance dynamically and achieve precise quality control without requiring multiple model instances

Inventive Principle:

Principle #23Feedback

Data Source

PatentUS12159438B2Substitutional input optimization for adaptive neural image compression with smooth quality control

Publication Date: 2024.12.03 TENCENT AMERICA LLC

AI summary

The present disclosure includes a method, apparatus, and non-transitory computer-readable medium for adaptive neural image compression by meta-learning. The method may include generating a substitute input image and a substitute target quality control parameter using an original input image and a target quality control parameter, wherein the substitute input image is a modified version of the original input image and the substitute target quality control parameter is a modified version of the target quality control parameter. The method may further include encoding the substitute input image, based on the substitute input image and the substitute target quality control parameter, using an encoding neural network, to generate a compressed representation of the substitute input image.