Unified Deep-Learning Filtering Across Intra- and Inter-Coded Pictures
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing loop filters based on deep learning in picture processing have limitations in performance optimization, and there is a need to reduce the cost of model parameter storage while improving filtering effects.
Innovation Solution
A general filtering model trained using deep learning is applied to perform filtering on pictures in both intra and inter encoding modes, reducing the storage space required for model parameters and enhancing the filtering effect by combining classification information with differential filtering.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If separate filtering models are built for different encoding modes (intra and inter), then filtering performance for each mode is optimized, but model parameter storage space increases significantly
Solution Approach 1:
The patent applies universality by designing a single filtering model that can handle both intra-frame and inter-frame encoding modes. The model incorporates classification information as an input feature, allowing one model to perform differential filtering for different modes without requiring separate specialized models, thus reducing storage space while maintaining filtering performance
Solution Approach 2:
The patent merges the functionality of multiple filtering models into one unified model. By combining intra-mode and inter-mode filtering capabilities into a single model structure that processes classification information alongside image data, the patent reduces the total model parameter storage space while preserving the filtering effects of mode-specific models
2Quantity of substance
If a single general filtering model is used for both intra and inter encoding modes, then model parameter storage space is reduced, but filtering performance may deteriorate due to lack of mode-specific optimization
Solution Approach 1:
The patent applies local quality by incorporating classification information as an additional input channel that provides mode-specific guidance to the filtering model. This allows the single model to adapt its filtering behavior locally based on the encoding mode (intra or inter), maintaining optimized filtering performance for each mode while using a single shared model architecture
Solution Approach 2:
The patent changes the input parameters of the filtering model by adding classification information as a new feature dimension. This additional parameter enables the model to distinguish between intra and inter encoding modes and adjust its filtering operations accordingly, maintaining high filtering performance while using a single unified model
Data Source
AI summary
In a picture filtering method, encoding modes corresponding to a plurality of picture areas in a picture are obtained. Encoding information of the picture is decoded. The encoding information includes classification information that is determined based on at least the encoding modes corresponding to the plurality of picture areas in the picture. The picture and the classification information are input into a general filtering model trained using deep learning. A filtered picture is obtained based on the general filtering model performing filtering on the picture based on the encoding information. The classification information includes at least one of first classification information and second classification information. The first classification information indicates a first encoding mode corresponding to one or more pixels in the picture. The second classification information indicates a second encoding mode corresponding to a preset size area in the picture.