Feature Map Scaling Intervals for Low-Complexity Entropy Coding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing adaptive quantization and latent scaling technologies in data coding, such as those used in JPEG AI, suffer from high computational complexity while maintaining coding performance.
Innovation Solution
Implement a method that sets non-overlapping threshold intervals for scaling features and probability distribution parameters, reducing the need for repeated scaling operations and thus lowering algorithm complexity without compromising coding performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If adaptive quantization and latent scaling technologies are used to improve coding performance, then coding performance is improved, but computational complexity increases
Solution Approach 1:
The feature map is divided into multiple non-overlapping intervals based on threshold values. Each interval corresponds to a specific scaling factor range. By segmenting the feature map into distinct intervals, the patent avoids repeated scaling operations across overlapping regions, thereby reducing computational complexity while maintaining coding performance through targeted scaling in each interval.
2Manufacturing precision
If multiple scaling operations are performed to maintain coding performance, then coding performance is maintained, but algorithm complexity increases
Solution Approach 1:
Instead of performing scaling operations on the entire feature map uniformly, the patent applies scaling selectively to specific intervals that meet certain criteria. By performing partial scaling only where necessary (in intervals with eigenvalues exceeding thresholds), the patent reduces the total number of scaling operations required, thereby lowering algorithm complexity while maintaining coding performance in the critical regions.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An encoding and decoding method and apparatus are provided. The method includes: obtaining a feature map of a to-be-encoded picture (1310); obtaining a probability distribution parameter map of the feature map (1320); obtaining a first matrix based on the probability distribution parameter map (1330); determining, from a plurality of intervals, a first interval corresponding to the first matrix, where the plurality of intervals do not overlap each other, and each interval corresponds to at least one scaling factor (1340); scaling, based on a first scaling factor corresponding to the first interval, the feature map and the probability distribution parameter map that correspond to the first matrix, to obtain a scaled probability distribution parameter map (1350); and performing entropy encoding on a scaled feature map based on the scaled probability distribution parameter map, and writing an entropy encoding result into a bitstream (1360). The method can reduce computational complexity of an adaptive quantization technology and a latent scaling technology while maintaining coding performance.