Face Authentication Vector Feature Masking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face authentication collation devices face challenges in accurately distinguishing between similar individuals due to high similarity scores from feature vectors, leading to potential false identifications, especially when features from occluded parts or backgrounds are not properly ignored.
Innovation Solution
A collation device that generates input and registered vectors by selecting specific features based on reference vectors, masking non-relevant features to calculate similarity, thereby reducing false positives by using first and second specification information to determine vector similarity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If feature vectors include all extracted features from image data, then the feature representation is comprehensive, but the degree of similarity between different persons becomes too high causing false identifications
Solution Approach 1:
The patent extracts and removes harmful features from the feature vector that cause false similarities between different persons. The feature vector generation unit selectively excludes features that are not discriminative or are harmful to collation accuracy, such as features from occluded regions or background elements, while retaining features that are specific to each person's identity.
Solution Approach 2:
The patent applies different quality standards to different features in the feature vector. Instead of treating all features uniformly, the system evaluates each feature's contribution to collation accuracy and selectively weights or excludes features based on their local quality and discriminative power, ensuring that only high-quality, person-specific features are used for similarity calculation.
2Reliability
If feature selection is performed independently on input and registered image data, then unnecessary features like occluded parts and backgrounds are ignored, but features indicating similar persons cannot be distinguished when they are similar
Solution Approach 1:
The patent merges the feature selection processes for input image data and registered image data into a unified feature vector generation unit. This integrated approach ensures that features are selected consistently for both datasets using the same discriminative criteria, enabling the system to distinguish between similar persons while maintaining collation accuracy. The unified selection process compares features across both datasets to identify discriminative patterns.
3Loss of information
If all features from image data are used for collation, then the feature representation is complete, but false positives increase when different persons have similar features
Solution Approach 1:
The feature vector generation unit extracts and removes harmful features that cause false similarities between different persons. By selectively excluding features that are not discriminative or are harmful to collation accuracy, the system maintains complete representation of person-specific features while eliminating features that lead to false positives.
Solution Approach 2:
The patent changes the parameters of the feature vector by selectively including or excluding specific features based on their discriminative power. The system dynamically adjusts the feature set used for collation, modifying the feature vector parameters to optimize the balance between information completeness and identification reliability by removing features that cause false similarities.
Data Source
AI summary
According to an embodiment, a collation device includes a hardware processor configured to: generate, based at least in part on input data, an input vector comprising input data features indicating features of the input data, the input data features comprising D number of features, D being an integer equal to or larger than two; andgenerate first specification information that specifies d selected features among the input data features of the input vector, based at least in part on a plurality of reference vectors and the input vector, the plurality of reference vectors each comprising reference features in the same form as the input vector, the reference features comprising the D number of features, d being an integer equal to or larger than one and smaller than D.