In-Loop Filter Model Selection for Low-Overhead Video Decoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video encoding and decoding systems face high complexity and additional bit overheads due to model selection schemes for in-loop filters, which determine whether to enable or disable models based on rate-distortion cost, leading to inefficient coding performance.
Innovation Solution
Implement a deep learning-based method for adaptive decision-making on in-loop filter network models, determining whether to use a target model or not for a current block, reducing complexity and avoiding bit overheads by using a preset selection network model to decide on output values for each candidate model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If rate-distortion cost calculation is used for model selection, then model performance is improved, but system complexity increases significantly
Solution Approach 1:
The patent changes the selection criterion from rate-distortion cost (which requires complex calculation) to a simplified parameter-based selection using a preset selection network model. The model directly outputs selected in-loop filter network models based on input features, avoiding complex cost calculations while maintaining selection accuracy.
Solution Approach 2:
The patent replaces the mechanical calculation process of rate-distortion cost with a neural network-based selection model. The preset selection network model learns optimal selection patterns during training and performs fast inference during encoding/decoding, substituting the computationally intensive mechanical calculation with a more efficient intelligent model.
2Adaptability or versatility
If model switch enable/disable information is signaled, then model adaptability is improved, but bit overhead increases
Solution Approach 1:
The patent extracts only the essential model selection information needed for reconstruction, removing redundant enable/disable signaling. The preset selection network model directly outputs the selected model indices, and only these compact indices are transmitted in the bitstream, eliminating unnecessary overhead while preserving adaptability.
Solution Approach 2:
The encoder and decoder both use the same preset selection network model, allowing the decoder to replicate the encoder's model selection decisions locally. This copying approach eliminates the need to transmit complex enable/disable control information, as the decoder can independently make the same selections using its copy of the selection model.
Data Source
AI summary
A decoding method includes: parsing bitstream to determine value of first syntax element identification information; determining, when the identification information indicates that preset selection network model is enabled to be used to perform model decision for current block, at least two output values according to preset selection network model. The at least two output values include first value corresponding to each candidate in-loop filter network model when in-loop filter network model is used for current block, and second value when in-loop filter network model is not used; determining, according to output values, target in-loop filter network model when in-loop filter network model is used for current block, or determining according to output values that in-loop filter network model is not used; and filtering, when in-loop filter network model is used for current block, current block using target in-loop filter network model, to obtain first reconstructed picture block of current block.