Adaptive In-Loop Filtering Model Selection for Video Decoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video encoding and decoding systems face complexity in selecting an in-loop filtering model with better performance due to the high computational cost of rate distortion cost calculations, leading to suboptimal filtering results for varying quality parameters.
Innovation Solution
Implement a method that uses a preset selection network model to adaptively select an in-loop filtering network model based on deep learning, allowing for model adaptive selection from candidate models to improve coding performance and decoding efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If rate distortion cost calculation is used to select filtering model, then filtering performance is improved, but computational complexity increases
Solution Approach 1:
The patent replaces the mechanical calculation process (rate distortion cost computation) with a neural network-based system. The selection network model learns optimal filtering model selection from training data, substituting complex mathematical calculations with learned patterns, thereby reducing computational complexity while maintaining or improving filtering performance.
Solution Approach 2:
The patent performs preliminary training of the selection network model offline using training sequences and reference software. This preliminary action transfers computational burden from real-time encoding/decoding to the offline training phase, allowing fast inference during actual video processing without sacrificing filtering performance.
2Productivity
If conventional filtering models are used, then encoding complexity is reduced, but video quality deteriorates
Solution Approach 1:
The patent introduces dynamic model selection by using a selection network that adaptsively chooses the most appropriate filtering model based on input characteristics. This dynamic approach replaces static conventional filtering with adaptive model selection, improving video quality while maintaining encoding efficiency through learned selection patterns.
Solution Approach 2:
The patent changes the selection criterion from rate distortion cost calculation to neural network output probabilities. By using the selection network's predicted probabilities to directly select filtering models, the system achieves better video quality through learned parameter selection while avoiding complex real-time calculations.
3Measurement precision
If multiple candidate filtering models are evaluated, then filtering accuracy is improved, but processing time increases
Solution Approach 1:
The patent uses the selection network model as a copy or approximation of the complex model evaluation process. Instead of actually evaluating multiple candidate filtering models through rate distortion calculations, the system uses the pre-trained selection network's predictions, which copy the essential selection logic without the computational overhead of full model evaluation.
Data Source
AI summary
Disclosed in the embodiments of the present application are an encoding method, a decoding method, a code stream, an encoder, a decoder and a storage medium. The decoding method comprises: parsing a code stream, and determining a value of first syntax element identification information; when the first syntax element identification information indicates that the current block is allowed to use a preset selection network model to perform model selection, determining the preset selection network model of the current block, and according to the preset selection network model, determining a loop filtering network model used by the current block; and performing filtering processing on the current block by using the loop filtering network model, so as to obtain a reconstructed image block of the current block.