Multistage Vector Quantization with Adaptive Candidate Search
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional multiple-stage vector quantization methods using N best search struggle to reduce encoding distortion while minimizing calculation, resulting in insufficient encoding performance.
Innovation Solution
A quantization apparatus and method employing tree search, where the number of candidates is determined in each stage based on previous stages' results, allowing for selective candidate retention and reduction to optimize quantization error reduction with minimal calculation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple-stage vector quantization uses N best search with N>1 to reduce encoding distortion, then encoding performance is improved, but calculation amount increases to N times
Solution Approach 1:
The patent applies dynamics by making the number of candidates N variable rather than fixed. The candidate number determination section dynamically adjusts N based on quantization error thresholds and stage progression, allowing the system to adapt between exploration (larger N) and exploitation (smaller N) throughout the quantization process, thereby resolving the contradiction between encoding distortion reduction and calculation amount
Solution Approach 2:
The patent changes the parameter of candidate number N throughout the quantization process. By modifying N based on quantization error criteria and stage number, the system optimizes the balance between search thoroughness (for lower distortion) and computational efficiency, directly addressing the technical contradiction
2Productivity
If only the quantization result with the smallest error is used in each stage to reduce calculation amount, then calculation efficiency is improved, but encoding distortion is not sufficiently reduced resulting in degraded quantization performance
Solution Approach 1:
The patent applies partial action by selectively retaining multiple candidates (N>1) only when necessary based on quantization error criteria, rather than always processing all candidates. This partial exploration maintains calculation efficiency while achieving sufficient encoding distortion reduction through staged candidate retention
Solution Approach 2:
The patent implements feedback by using quantization error results from each stage to determine the number of candidates for subsequent stages. The candidate number determination section continuously adjusts N based on feedback from quantization performance, optimizing the balance between calculation efficiency and encoding distortion
Data Source
AI summary
A quantization device and quantization method are provided that reduce coding distortion with a small degree of calculation and achieve adequate coding performance thereby. A multistage vector quantization unit treats a number of candidates N that are designated prior to operation in the first-stage vector quantization unit, decrements the number of candidates by one beginning with the second-stage vector quantization unit and continuing with each stage thereafter. If the number of candidates is three or less, the multistage vector quantization unit assesses the quantization distortion at each stage, treating the number of candidates at the following stage as a predetermined value P if the quantization distortion is greater than a prescribed threshold, and treating the number of candidates at the following stage as a value Q that is less than the predetermined value P if the quantization distortion is less than or equal to the predetermined threshold.