Neural Codec Entropy Coding With Channel-Wise PMF Grouping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural codecs face challenges in efficiently compressing and restoring data while minimizing hardware requirements and power consumption, particularly due to the large memory and computational overhead of entropy coding tables.
Innovation Solution
A neural codec that encodes input data into a first latent vector, groups channels, predicts probability mass functions (PMFs) for entropy encoding, and uses a look-up table (LUT) based on Huffman coding to generate a bitstream, optimizing channel-wise grouping and reducing memory usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If entropy coding tables are used to improve compression efficiency, then compression efficiency is improved, but hardware area and power consumption increase
Solution Approach 1:
The patent extracts the essential function of entropy coding by replacing traditional large lookup tables with a neural network-based probability estimation system. The neural network models the probability distribution of quantized coefficients and generates probability mass functions dynamically, eliminating the need for storing extensive pre-computed entropy coding tables in hardware.
Solution Approach 2:
The patent substitutes the mechanical/storage-based entropy coding system with a computational neural network system. Instead of using hardware lookup tables to store and retrieve entropy codes, the system uses a neural network to compute probability distributions and generate codes algorithmically, trading hardware resources for computational processing.
2Productivity
If entropy coding tables are used to improve compression efficiency, then compression efficiency is improved, but power consumption increases
Solution Approach 1:
The patent removes the power-intensive lookup table storage and access mechanisms by replacing them with a neural network that computes probability distributions in real-time. This extraction of the traditional entropy coding table approach significantly reduces the hardware resources required and consequently lowers power consumption during the encoding process.
Solution Approach 2:
The patent replaces the energy-consuming hardware-based lookup table system with a software-based neural network computation system. The neural network processes data through computational layers to estimate probabilities and generate codes, which consumes less power than the hardware lookup table access and management operations.
3Area of stationary object
If channel-wise grouping is applied to reduce memory usage, then memory usage is reduced, but compression efficiency may deteriorate
Solution Approach 1:
The patent applies local quality by grouping channels based on the statistical properties and probability distributions of their quantized coefficients. Channels with similar probability characteristics are grouped together, allowing the system to use a single neural network model for each group. This localized approach reduces memory requirements for storing multiple distinct models while maintaining compression efficiency through tailored probability estimation for each channel group.
Data Source
AI summary
An encoding method of a neural codec performed by at least one processor includes encoding input data to a first latent vector of a quantized representation; generating, after encoding the input data, a bitstream corresponding to the input data by channel-wise grouping and entropy encoding the first latent vector; and transmitting the bitstream.