Hyperprior Probability Estimation for Multi-Bit-Rate Image Coding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In multi-bit-rate scenarios, the numerical range of image features varies greatly due to different quantization steps, making it difficult to train a probability distribution estimation network with good performance, thus affecting encoding and decoding performance.
Innovation Solution
Estimate the probability distribution of unquantized image features using a hyperprior feature via a probability distribution estimation network, and then quantize the result to obtain a stable numerical range, simplifying network training and improving performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If different quantization steps are used to match different bit rates, then bit rate adaptability is improved, but the numerical range of image features varies greatly, making network training difficult and unstable
Solution Approach 1:
The patent segments the probability distribution estimation into two independent parts: (1) estimating the probability distribution of unquantized image features using a hyperprior feature, and (2) quantizing the estimated probability distribution parameters separately. This segmentation allows the estimation network to operate on stable, unquantized features while still supporting multiple bit rates through separate quantization steps.
Solution Approach 2:
The patent introduces the hyperprior feature as an intermediary to estimate the probability distribution parameters. Instead of directly quantizing image features before probability estimation, the hyperprior feature serves as a stable intermediate representation that captures the statistical properties of the unquantized features, enabling reliable training across different bit rates.
2Productivity
If quantization operations are performed to match bit rates, then encoding efficiency is improved, but encoding and decoding performance deteriorates due to numerical range variations
Solution Approach 1:
The patent performs preliminary probability distribution estimation on unquantized image features before quantization. By estimating the probability distribution parameters (mean and variance) of the unquantized features first, the system maintains the numerical stability needed for good reconstruction performance while still applying quantization for encoding efficiency.
Solution Approach 2:
The patent changes the parameter representation by working with probability distribution parameters (mean and variance) rather than raw image feature values. This parameter transformation allows the system to handle different quantization steps more effectively, as the probability parameters can be quantized separately while preserving the underlying feature distribution characteristics.
Data Source
AI summary
When probability estimation is performed in encoding and decoding processes, probability distribution of an unquantized image feature is estimated based on a hyperprior feature of the unquantized image feature via a first probability distribution estimation network, and then probability distribution of a quantized image feature is obtained through quantization. Alternatively, probability distribution of a quantized image feature is directly estimated based on a hyperprior feature of an unquantized image feature via a second probability distribution estimation network.