Neural Network Non-Linear Filter for Image Coding

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional image encoding methods, such as those in the HEVC standard, face inefficiencies in predicting images and filtering, leading to increased code size and potential image quality deterioration due to the use of linear filters and fixed filter coefficients.

Innovation Solution

The implementation of a neural network-based non-linear filter process for intra prediction, inter prediction, and in-loop filtering, which determines optimal parameters for each block based on pixel variance and distribution, allowing for more accurate prediction and filtering.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If linear filters and fixed filter coefficients are used for intra prediction and in-loop filtering, then device complexity is reduced and ease of manufacture is improved, but image quality deteriorates and encoding efficiency decreases

Engineering Contradiction:
Improveease of implementationVSAvoidimage quality
Core Design Contradiction:
Ease of manufactureVSManufacturing precision

Solution Approach 1:

The patent changes the parameters of the filtering process by introducing non-linear filters with adaptive coefficients. Specifically, filter coefficients are no longer fixed but are adaptively determined based on local image characteristics such as gradient information and variance, allowing the filtering strength and behavior to vary across different regions of the image, thereby improving image quality while maintaining reasonable implementation complexity

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent introduces dynamic filtering by making filter coefficients adaptive rather than static. The filtering process dynamically adjusts its behavior based on local image content, using metrics like gradient magnitude and variance to determine appropriate filter strengths for different blocks, transforming a static linear filtering approach into a dynamic non-linear approach that responds to image characteristics

Inventive Principle:
Principle #15Dynamics

2Device complexity

If conventional linear filtering is used, then computational complexity is reduced, but prediction accuracy decreases leading to increased code size

Engineering Contradiction:
Improvecomputational complexityVSAvoidprediction error
Core Design Contradiction:
Device complexityVSLoss of information

Solution Approach 1:

The patent changes the mathematical nature of the filtering operation from linear to non-linear, and transforms fixed parameters into adaptive parameters. By using non-linear functions of pixel values and adaptive coefficients derived from local image statistics, the filtering achieves better prediction accuracy without requiring excessive computational resources, as the complexity is localized to coefficient calculation rather than the entire filtering process

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The filtering system performs self-adjustment by automatically determining appropriate filter coefficients based on local image characteristics. The process uses locally computed metrics such as gradient information and variance to self-determine the optimal filtering parameters for each block, eliminating the need for external manual tuning or complex global optimization while improving prediction accuracy

Inventive Principle:
Principle #25Self-service

3Productivity

If fixed filter coefficients are used, then encoding speed is improved, but prediction precision deteriorates resulting in larger code size

Engineering Contradiction:
Improveencoding speedVSAvoidprediction precision
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent transforms fixed filter coefficients into adaptive parameters that are calculated based on local image characteristics. By deriving coefficients from local gradient information and variance, the system achieves better prediction precision while maintaining encoding efficiency, as the coefficient calculation is performed using simple local statistics rather than complex global optimization

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent divides the image into blocks and applies different filtering parameters to each block based on its local characteristics. By segmenting the image and computing adaptive coefficients for each segment independently using local statistics, the system achieves high prediction precision without requiring complex inter-block dependencies, thus maintaining encoding speed

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP3310058B1Image coding method, image decoding method, image coding device and image decoding device
Publication Date: 2023.02.22 PANASONIC INTELLECTUAL PROPERTY MANAGEMENT CO LTD
  • EP3310058B1 patent drawingFigure 1
  • EP3310058B1 patent drawingFigure 2
  • EP3310058B1 patent drawingFigure 3

AI summary

The image encoding method includes: encoding an image including a block by performing, in at least one of intra prediction, inter prediction, and an in-loop filter, a non-linear process by which the input-output relationship becomes non-linear (S11); and encoding an operation parameter of a non-linear filter to be used in the non-linear process (S12).