Neural Network Video Encoder Filtering Residual Redundancy
Solution Overview
Problem
Traditional video compression schemes have low compression efficiency due to limitations in motion estimation, motion compensation, and residual compression.
Innovation Solution
A method for encoding and decoding video that involves determining side information correlated with a current coding unit, filtering relevant information using a preset network model and side information, and inputting the filtered information into a subsequent coding module to generate a bitstream.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional motion estimation, motion compensation and residual compression are used, then the video compression process is simple and easy to implement, but the compression efficiency is low
Solution Approach 1:
The patent introduces a neural network model as an intermediary component between traditional video compression modules. The neural network processes residual information and generates refined prediction values, acting as a mediator that enhances the performance of traditional motion estimation and compensation while removing redundant information more effectively.
Solution Approach 2:
The patent changes the parameter representation by using learned filters and neural network parameters instead of fixed traditional compression parameters. The system learns optimal filtering parameters from data, transforming the compression process from using predetermined parameters to using adaptive, learned parameters that improve compression efficiency.
2Productivity
If more redundant information is removed to improve compression efficiency, then the compression ratio increases, but the accuracy of prediction values may deteriorate
Solution Approach 1:
The patent implements a feedback mechanism where the neural network model processes residual information and feeds back refined prediction values to improve the reconstruction quality. This feedback loop ensures that even when redundant information is aggressively removed, the prediction accuracy is maintained or improved through iterative refinement.
Solution Approach 2:
The patent performs preliminary filtering and neural network processing on residual information before final reconstruction. By preparing and refining prediction values in advance through learned filters and neural networks, the system ensures high prediction accuracy is achieved before the actual compression and reconstruction steps.
3Loss of information
If a preset network model is used to filter relevant information, then redundant information is effectively removed, but the computational complexity increases
Solution Approach 1:
The patent segments the video compression process into distinct functional modules: traditional motion estimation, residual calculation, neural network filtering, and final reconstruction. By dividing the complex process into manageable segments, each with a specific function, the system can apply sophisticated neural network filtering without overwhelming computational complexity.
Solution Approach 2:
The patent applies partial filtering through the neural network model, focusing computational resources on processing only the residual information that contains redundant patterns. Rather than processing the entire video data through complex neural networks, the system applies the model selectively to where it provides the most benefit, achieving efficient redundant information removal with controlled computational cost.
Data Source
AI summary
A video encoding method is applicable to an encoder, and comprises: determining side information having a correlation with a current coding unit; filtering first relevant information to be coded of the current coding unit by using a preset network model and the side information, to obtain second relevant information to be coded; and inputting the second relevant information to be coded into a subsequent coding module for encoding, to obtain a bitstream.