Decoder In-Loop Filtering Using QP-Adaptive Neural Network Flags
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video coding technologies face challenges with high complexity and overhead in hardware implementation due to multi-model intra switchable solutions, and inflexibility in model selection for in-loop filtering, leading to poor performance in chroma colour components.
Innovation Solution
A method that determines syntax element flag information to selectively apply a preset network model based on block quantization parameters for luma and chroma colour components, reducing unnecessary inputs and optimizing model usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multi-model intra switchable solution is used for in-loop filtering, then filtering performance can be adjusted according to local details, but hardware implementation complexity and overhead increase significantly
Solution Approach 1:
The patent uses a single neural network model that can process multiple colour components (luma and chroma) by taking component type and quantization parameter information as inputs. This universal model replaces the need for multiple separate models, reducing hardware complexity while maintaining the ability to adapt to different filtering requirements through parameter variations rather than model switching.
Solution Approach 2:
The patent changes the approach from model switching to parameter-based adaptation. By varying quantization parameters and component type inputs to the same neural network model, the system achieves different filtering behaviors without requiring multiple models, thus reducing hardware complexity while preserving adaptive filtering performance.
2Device complexity
If intra non-switchable solution with only two neural network models is used, then complexity is reduced and model generalization capability is improved, but selection during loop filtering becomes inflexible due to quantization parameters
Solution Approach 1:
The patent enables flexible model selection by using quantization parameter information as dynamic inputs to the neural network model. Different quantization parameters guide the model to produce different filtering results, providing adaptability across various picture types and quality requirements without requiring multiple separate models or complex switching logic.
Solution Approach 2:
The patent introduces dynamic adaptability by making the neural network model's behavior dependent on runtime parameters such as quantization values and component type. This allows the single model to dynamically adjust its filtering characteristics based on the specific picture being processed, achieving flexibility without multiple static models.
3Device complexity
If a single neural network model is used for both luma and chroma colour components, then hardware complexity is reduced, but chroma colour component filtering performance deteriorates
Solution Approach 1:
The patent applies local quality by treating luma and chroma components differently within the same neural network model. By incorporating component type information and specific quantization parameters for each component, the model can apply different filtering strategies locally optimized for each colour component type, maintaining high performance for both luma and chroma without requiring separate models.
Solution Approach 2:
The patent uses different quantization parameter sets for luma and chroma components as inputs to the neural network model. This parameter differentiation allows the single model to adapt its processing to the specific characteristics of each colour component, ensuring optimal filtering performance for chroma while maintaining hardware efficiency.
Data Source
AI summary
A coding method and decoder are disclosed. The method includes the following. A bitstream is parsed to determine first syntax element flag information of a to-be-filtered component of a current picture. When the first syntax element flag information indicates that there is a to-be-filtered component of a partition block in the current picture that is allowed to be filtered by using a preset network model, the bitstream is parsed to determine second syntax element flag information of a to-be-filtered component of a current block. When the second syntax element flag information indicates that the to-be-filtered component of the current block is to be filtered by using the preset network model, block quantization parameter information of the current block is determined, where the block quantization parameter information at least includes a block quantization parameter value of a first colour component and a block quantization parameter value of a second colour component.