Neural Visual Coding Quantization for Device-Interoperable Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image/video compression technologies face challenges in achieving efficient coding and interoperability across different devices, particularly in neural network-based methods, due to issues like quantization loss and overflow during convolution operations, which affect coding efficiency and device compatibility.
Innovation Solution
Implementing quantization and scaling processes for visual data processing, along with neural network models characterized by specific parameters such as channel and layer thresholds, group convolutions, and kernel sizes, to ensure interoperability and improve coding efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If floating-point calculations are used in neural network-based image compression, then calculation precision is maintained, but device interoperability deteriorates and overflow errors occur during convolution operations
Solution Approach 1:
The patent changes the numerical representation parameter from floating-point to fixed-point format. This parameter change resolves the contradiction by maintaining sufficient calculation precision while ensuring consistent behavior across different devices and avoiding overflow errors during convolution operations.
Solution Approach 2:
Instead of using the conventional floating-point format that causes interoperability issues, the patent inverts the approach by using fixed-point format. This inversion solves the device interoperability problem while maintaining adequate precision for image compression tasks.
2Reliability
If the number of channels, layers, and kernel sizes are reduced in the neural network model, then device interoperability is improved, but compression performance deteriorates
Solution Approach 1:
The patent applies parameter changes by standardizing architectural parameters (channels, layers, kernel sizes) to interoperable values. This allows the model to run consistently across devices while maintaining effective compression performance through optimized fixed-point operations.
3Reliability
If quantization process is applied to convert floating-point to fixed-point, then device interoperability is improved, but quantization loss increases
Solution Approach 1:
The patent applies preliminary scaling and normalization before quantization to minimize information loss. By preparing the data in advance with appropriate scaling factors, the quantization process achieves both device interoperability and reduced quantization loss.
Solution Approach 2:
The patent uses feedback mechanisms during training to optimize the quantization process. The model learns to compensate for quantization effects, allowing it to maintain compression performance while benefiting from the interoperability improvements of fixed-point arithmetic.
Data Source
AI summary
Embodiments of the present disclosure provide a solution for visual data processing. A method for visual data processing is proposed. The method comprises: performing, for a conversion between visual data and a bitstream of the visual data, a quantization process on a dataset comprising at least one of: input visual data of a neural network model used for the conversion, or a parameter of the neural network model; and performing the conversion based on the quantization process.