Hash-Based Face Recognition Using Statistical Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Biometric identification systems face challenges in efficiently and quickly recognizing human faces due to variations in shape and appearance caused by factors like emotion, expression, age, and lighting conditions, making it difficult to create an exhaustive library of human faces and their many variations.
Innovation Solution
A system and method that uses statistical models to handle variations in an object class, combined with feature points to correlate features across different images, by aligning a test image with a canonical sample, applying patch patterns, computing binary features, and using hash keys to identify the highest probability match in a hash table.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If an exhaustive library of human faces and their variations is created, then recognition accuracy is improved, but device complexity and time required increase significantly
Solution Approach 1:
The patent extracts only the essential facial features and structural characteristics from the vast space of possible face variations. By identifying and storing only these key features rather than complete face images, the system achieves accurate recognition without requiring an exhaustive library of all possible face variations.
Solution Approach 2:
The patent creates a simplified representation or copy of facial features using mathematical models and feature vectors. These copies capture the essential identifying characteristics while discarding redundant information, allowing for efficient storage and comparison without needing the actual complete face images.
2Measurement precision
If an exhaustive library of human faces is created, then recognition accuracy is improved, but processing time increases
Solution Approach 1:
The patent extracts only the essential facial features and structural characteristics from the vast space of possible face variations. By identifying and storing only these key features rather than complete face images, the system achieves accurate recognition without requiring an exhaustive library of all possible face variations.
Solution Approach 2:
The patent transforms facial images into parameterized representations such as feature vectors, eigenfaces, or statistical models. This parameter transformation converts the recognition problem from comparing entire images to comparing compact parameter sets, significantly reducing processing time while maintaining recognition accuracy.
3Adaptability or versatility
If statistical models are used to handle face variations, then adaptability is improved, but device complexity increases
Solution Approach 1:
The patent applies different levels of statistical modeling to different aspects of face recognition. Instead of using complex models for all features, the system uses simple feature extraction for stable characteristics and more complex models only where necessary, such as for handling expression variations or lighting conditions.
Solution Approach 2:
The patent employs dynamic statistical models that can adapt to variations in facial appearance through parameters like mean faces, covariance matrices, and probability distributions. These models dynamically adjust to accommodate different expressions, ages, and lighting conditions without requiring separate models for each variation.
Data Source
AI summary
In a face recognition system, overlapping patches are defined on a canonical face. Random clusters of pixel pairs are defined within each patch, and binary features are determined for each pixel pair by comparing their respective feature values. An inverted index hash table is constructed of the binary features. Similar binary features are then determined on a library of registrable samples of identified faces. A log probability of each registrable sample generating a binary feature from a corresponding cluster of pixel pairs at each specific patch location is determined and stored in the hash table. In a search phase, similar binary features are determined, and a hash key is determined for each binary feature. The log probabilities for each identity found in the hash table are summed for all clusters of pixel pairs and locations and sorted to find the high probability match.