Block-Based Neural Decoding Using Probability Models and Sub-Blocks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network-based coding and decoding technologies suffer from poor coding performance, poor decoding performance, and high complexity, particularly in video processing.
Innovation Solution
A block-based coding and decoding solution that utilizes probability distribution parameters and mean features to improve coding and decoding performance, reducing complexity and memory usage while maintaining high-quality image reconstruction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If neural network-based coding and decoding is used, then coding performance and decoding performance can be improved, but complexity increases significantly
Solution Approach 1:
The patent divides the feature block into multiple stage sub-blocks and processes them in stages. The analysis transform network and synthesis transform network are applied hierarchically to different levels of the image block structure, breaking down the complex neural network processing into manageable segments that reduce overall computational complexity while maintaining coding performance.
Solution Approach 2:
The patent introduces probability distribution parameters that dynamically adapt to different image blocks and stages. The coding and decoding processes use dynamic probability models that adjust based on the specific characteristics of each sub-block, allowing the system to optimize performance for different content types without requiring a single complex fixed architecture.
2Manufacturing precision
If neural network-based coding and decoding is used, then coding performance and decoding performance can be improved, but memory usage increases
Solution Approach 1:
The patent segments the feature block into multiple stage sub-blocks that can be processed independently. This segmentation allows memory to be allocated and released in stages rather than requiring all features to be held in memory simultaneously, significantly reducing peak memory usage while maintaining the ability to reconstruct the full image block.
Solution Approach 2:
The patent performs analysis transform and extracts probability distribution parameters in advance before the actual decoding process. This preliminary processing allows the system to prepare and store only essential parameters rather than full feature representations, reducing memory requirements for the main decoding operation.
3Device complexity
If traditional coding methods are used, then complexity is low, but coding performance and decoding performance are poor
Solution Approach 1:
The patent introduces probability distribution parameters as intermediary representations between the raw image data and the final coded output. These parameters serve as a bridge that captures essential statistical properties of image blocks, enabling efficient coding with simpler methods while maintaining performance that would otherwise require complex neural networks.
Solution Approach 2:
The patent transforms image features into probability distribution parameters through the analysis transform network. This parameter transformation converts complex spatial and temporal correlations into probabilistic representations that can be efficiently coded using simplified methods, achieving high coding performance with reduced complexity.
4Productivity
If block-based processing is used, then parallel decoding speed is improved, but processing overhead increases
Solution Approach 1:
The patent divides the image block into multiple stage sub-blocks that can be processed in parallel. Each sub-block undergoes independent analysis transform, probability parameter extraction, and synthesis transform, allowing simultaneous processing of multiple regions. This segmentation enables efficient parallelization that increases decoding speed while the modular structure keeps processing overhead manageable.
Data Source
Figure 1~2
Figure 3~4
Figure 5~6
AI summary
The present disclosure provides decoding methods and apparatuses, coding methods and apparatuses, devices and media thereof, the decoding method includes: obtaining, by decoding a first bitstream of a current image block, a coefficient hyperparameter feature of each of stage sub-blocks in the current image block; for each of the stage sub-blocks, determining a probability distribution parameter based on the coefficient hyperparameter feature of the stage sub-block, and obtaining, by decoding a second bitstream of the current image block based on the probability distribution parameter, a residual feature of the stage sub-block; determining a reconstructed feature of the stage sub-block based on the residual feature of the stage sub-block and a mean feature of the stage sub-block; and determining a reconstructed image block corresponding to the current image block based on respective reconstructed features of the stage sub-blocks. Through the technical solutions of the present disclosure, coding performance and decoding performance can be improved.