Relative Entropy Quantization for Low-Artefact Signal Coding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Quantization of audio or video signals often results in perceptible noise and artefacts, such as 'birdies,' especially at moderate bit rates, which existing methods struggle to mitigate effectively without increasing bandwidth or storage requirements.
Innovation Solution
A method and device that minimize quantization error by determining a partition into quantization cells and a reconstruction probability distribution, ensuring the difference between the source signal and reconstructed signal distributions is constrained, thereby reducing artefacts and noise.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If quantization is applied to reduce bandwidth and storage requirements, then transmission efficiency is improved, but perceptible quantization noise and artefacts occur in the reconstructed signal
Solution Approach 1:
The patent changes the parameter of quantization resolution dynamically. Instead of using uniform quantization steps, the system optimizes quantizer resolution to minimize average distortion given a fixed rate, leading to variable quantization resolution where finer steps are used in regions of high signal importance and coarser steps in less critical regions.
Solution Approach 2:
The patent applies different quantization resolutions to different parts of the signal based on local characteristics. In transform-based quantization of audio signals, the system applies variable resolution quantization where the quantization step size varies across frequency components and time frames, allowing high-fidelity representation of important signal components while using coarser quantization for less critical components.
2Measurement precision
If quantizer resolution is optimized to minimize average distortion at fixed rate, then signal quality is improved, but variable quantization resolution increases device complexity
Solution Approach 1:
The patent performs preliminary optimization of the quantizer design before actual signal encoding. The system pre-calculates optimal quantization parameters, resolution levels, and step sizes based on signal statistics and distortion criteria, storing these parameters for use during encoding. This preliminary action reduces the computational complexity during real-time operation.
Solution Approach 2:
The patent implements dynamic quantization where the quantizer resolution and step sizes are adjusted based on signal characteristics. The system adapts the quantization parameters dynamically according to the local signal properties, such as signal amplitude, frequency content, and temporal variations, allowing optimal signal quality while managing complexity through adaptive rather than purely variable approaches.
3Object-affected harmful factors
If dithering noise is added to improve audible impression, then perceptibility of artefacts is reduced, but mean squared error increases and theoretical upper bound on distribution difference is unbounded
Solution Approach 1:
The patent converts the harmful effect of quantization noise into a beneficial effect by deliberately adding dithering noise before quantization. This stochastic noise randomizes the quantization error, preventing the formation of correlated artefacts and distortion products. The dithering noise masks the quantization errors in a way that makes them less perceptible to human perception, even though it increases the mean squared error.
Solution Approach 2:
The patent introduces dithering noise as an intermediary element between the original signal and the quantization process. This intermediate noise component acts as a mediator that modifies the quantization error characteristics, transforming deterministic quantization errors into stochastic errors that are less perceptible. The dithering noise serves as a buffer that decouples the direct relationship between signal and quantization error.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Methods and devices for encoding and decoding are provided. A source signal value is encoded by a quantization index determined using a partition into quantization cells. Decoding of the quantization index takes place by sampling a reconstruction probability distribution, thereby obtaining a reconstructed signal value, such that the reconstructed signal value lies in the same quantization cell as the source signal value. In one embodiment, encoding and decoding are such that their succession preserves the source signal distribution. In another embodiment, the partition and the reconstruction probability distribution are determined in such manner that the quantization error is minimized subject to a constraint on the relative entropy between the source signal and the reconstructed signal.