Rank-Based Score Normalization for Biometric Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current score normalization techniques in biometric systems face challenges in effectively handling variations in pose, illumination, and other conditions, leading to inconsistent matching scores and reduced performance, especially in unimodal systems with multiple samples per subject.
Innovation Solution
A rank-based score normalization framework that partitions matching scores into subsets, normalizes each subset independently, and utilizes gallery-based information to align score distributions, allowing for improved discriminability and integration of scores across multiple samples.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If global score normalization is applied to handle variations in pose and illumination, then score comparability is improved, but detection and identification rates deteriorate due to inconsistent matching scores
Solution Approach 1:
The patent partitions the gallery into multiple subsets based on similarity scores, and applies separate normalization to each subset rather than using a single global normalization. This segmentation allows the system to handle different score distributions in different regions, improving both adaptability to variations and reliability of detection/identification rates
Solution Approach 2:
The patent applies different normalization strategies to different local regions (subsets) of the score space. By treating high-scoring and low-scoring regions differently with subset-specific normalization parameters, the system achieves local optimization that improves overall system performance while maintaining score comparability
2Quantity of substance
If multiple samples per subject are used in the gallery, then information availability is improved, but score distribution consistency deteriorates leading to reduced matching performance
Solution Approach 1:
The patent segments the gallery samples into subsets based on their similarity scores, and applies separate normalization to each subset. This allows the system to utilize multiple samples per subject while maintaining score distribution consistency within each subset, resolving the contradiction between information availability and distribution stability
Solution Approach 2:
The patent changes the normalization parameters (mean and standard deviation) for each subset separately rather than using fixed global parameters. This adaptive parameter adjustment allows the system to handle the varied score distributions introduced by multiple samples per subject, maintaining consistency while utilizing abundant information
Data Source
AI summary
A rank-based score normalization framework that partitions matching scores into subsets and normalize each subset independently. Methods include implementing two versions of the framework: (i) using gallery-based information (i.e., gallery versus galleryscores), and (ii) updating available information in an online fashion. The methods improve the detection and identification rate from 20:90% up to 35:77% for Z-score and from 25:47% up to 30:29% for W-score.