Signaling Neural Network Post-Filter Parameters in Video Coding

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current video coding standards, such as ITU-T H.264 and H.265, lack efficient methods for signaling neural network post-filter parameter information, which limits the ability to effectively utilize advanced post-processing techniques for improving video quality.

Innovation Solution

The proposed solution involves signaling neural network post-filter parameter information through specific syntax elements in the video coding process, including indicating the presence of matrices for input tensors and conditionally signaling the bit depth of chroma sample values, to enable more effective post-filtering operations.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If neural network post-filter parameter information is not signaled in current video coding standards, then the coding structure remains simple and compatible with existing standards, but the ability to utilize advanced post-processing techniques for improving video quality is limited

Engineering Contradiction:
Improvevideo qualityVSAvoidcoding structure
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent segments the neural network parameter information into distinct syntax elements that are separately signaled in the bitstream. This includes dividing parameters into different categories (e.g., tensor dimensions, activation functions, quantization parameters) that can be independently controlled and transmitted, allowing the system to adopt advanced post-processing capabilities while maintaining a structured and manageable coding framework

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces dynamic signaling mechanisms where neural network parameters can be adaptively selected and transmitted based on content requirements. The system allows for conditional signaling of syntax elements and supports different precision levels for different parameter types, enabling the coding structure to dynamically adjust its complexity based on the specific post-processing needs without requiring a completely new rigid framework

Inventive Principle:
Principle #15Dynamics

2Manufacturing precision

If comprehensive neural network parameters are always signaled, then post-filter performance is maximized, but bitstream overhead increases

Engineering Contradiction:
Improvepost-filter performanceVSAvoidbitstream overhead
Core Design Contradiction:
Manufacturing precisionVSQuantity of substance

Solution Approach 1:

The patent applies local quality by signaling neural network parameters with different precision levels based on their specific requirements. Not all parameters are transmitted with the same bit depth - critical parameters may use higher precision while less sensitive parameters use lower precision. This selective precision approach ensures adequate post-filter performance for important parameters while reducing overall bitstream overhead

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent enables dynamic parameter changes by allowing the encoder to select from multiple predefined neural network configurations and signal only the necessary identifiers rather than complete parameter sets. The system supports switching between different neural network models or configurations based on content characteristics, transmitting only the relevant parameter subsets needed for each specific case, thereby optimizing the balance between performance and overhead

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4447451A1Systems and methods for signaling neural network post-filter tensor purpose and order information in video coding
Publication Date: 2024.10.16 SHARP KK
  • EP4447451A1 patent drawingFigure 1
  • EP4447451A1 patent drawingFigure 2
  • EP4447451A1 patent drawingFigure 3

AI summary

A device may be configured to perform filtering based on information included in a neural network post-filter characteristics message. In one example, the neural network post-filter characteristics message includes a syntax element indicating matrices which are present for an input tensor of a neural network post-filter. In one example, the neural network post-filter characteristics message conditionally includes a syntax element indicating a bit depth of chroma sample values, in cases where chroma matrices are present.