Codebook Quantization Alignment for DNN Compression

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing data compression methods for video and Deep Neural Networks (DNNs) face inefficiencies in compression efficiency, particularly when using codebook-based quantization with entropy coding techniques like CABAC, leading to suboptimal performance due to misalignment between quantization outputs and entropy coding inputs.

Innovation Solution

The method encodes and decodes data by adapting the output of quantization to align with entropy coding, using a flag to select between uniform and codebook-based quantization, and modifying the codebook and indices to center coded values around zero, optimizing the use of regular CABAC bins and omitting the sign flag when necessary.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If codebook-based quantization is used with entropy coding, then compression efficiency is improved, but misalignment between quantization outputs and entropy coding inputs occurs

Engineering Contradiction:

Improvecompression efficiencyVSAvoidalignment between quantization and entropy coding

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The patent modifies the codebook values and indices by centering them around zero, changing the parameter distribution to match the expectations of CABAC entropy coding. This parameter transformation resolves the misalignment issue while preserving compression efficiency gains from codebook-based quantization.

Inventive Principle:

Principle #35Parameter changes

Solution Approach 2:

The patent introduces a flag field as an intermediary to signal the type of quantization used (uniform or codebook-based). This mediator allows the entropy coding process to adapt its behavior based on the quantization method, ensuring proper alignment between quantization outputs and entropy coding inputs.

Inventive Principle:

Principle #24Intermediary (Mediator)

2Adaptability or versatility

If uniform quantization is used, then compatibility with existing standards is maintained, but compression efficiency is reduced

Engineering Contradiction:

Improvecompatibility with existing standardsVSAvoidcompression efficiency

Core Design Contradiction:

Adaptability or versatilityVSProductivity

Solution Approach 1:

The patent makes the quantization method dynamic by using a flag to select between uniform and codebook-based quantization. This allows the system to adapt to different scenarios: using uniform quantization for compatibility with existing standards like MPEG NNR, and codebook-based quantization when higher compression efficiency is needed.

Inventive Principle:

Principle #15Dynamics

3Ease of operation

If sign flag is encoded during binarizing, then all quantized data is handled uniformly, but additional bits are transmitted unnecessarily

Engineering Contradiction:

Improveuniform handling of quantized dataVSAvoidnumber of bits transmitted

Core Design Contradiction:

Ease of operationVSQuantity of substance

Solution Approach 1:

The patent applies partial action by encoding the sign flag only when necessary (for uniform quantization) and omitting it when codebook-based quantization is used. This selective encoding avoids transmitting unnecessary bits while maintaining ease of operation through conditional logic based on the quantization type flag.

Inventive Principle:

Principle #16Partial or excessive action

Data Source

PatentUS20230064234A1Systems and methods for encoding a deep neural network

Publication Date: 2023.03.02 INTERDIGITAL MADISON PATENT HLDG

AI summary

The disclosure relates to a method for compression including codebook-based quantization of a data set and corresponding decompression method, signal; bitstream, and encoder and/or decoder device.