High Resolution Face Matching via Segmented Part Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current facial recognition systems provide only partial information about the relationship between faces, necessitating a higher resolution face matching scheme.
Innovation Solution
A method for high-resolution face matching that involves obtaining and aligning visual representations of faces, virtually segmenting them into parts, masking and matching these parts, and generating relevancy scores to provide detailed match results, utilizing outputs from intermediate layers of a convolutional neural network for enhanced feature vector analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If a global score is provided for face matching, then the system is simple and fast, but the information provided is partial and lacks detail
Solution Approach 1:
The patent applies segmentation by dividing the face matching process into multiple independent part-level matchers, each responsible for specific facial regions (eyes, nose, mouth, etc.). This allows the system to provide detailed information about which specific parts match or don't match, rather than just a global score, thereby reducing information loss while maintaining manageable system complexity through modular design.
2Measurement precision
If part-level face matching is implemented, then detailed match information is provided, but the computational complexity increases
Solution Approach 1:
The face is segmented into multiple distinct parts (eyes, nose, mouth, cheeks, etc.), and separate matchers are trained for each part using specialized loss functions. This segmentation enables precise measurement of matching quality at the part level while managing computational complexity through efficient architecture design and selective application of complex loss functions only where needed.
Solution Approach 2:
Different loss functions and matching strategies are applied to different facial parts based on their specific characteristics. For example, eyes may use different matching criteria than the mouth, allowing each region to be evaluated with the most appropriate precision metrics while avoiding uniform computational overhead across all regions.
3Reliability
If multiple loss functions are used for training, then matching accuracy improves, but training complexity increases
Solution Approach 1:
Different loss functions are selectively applied to different facial parts based on their specific matching requirements. For instance, certain parts may use classification loss while others use regression loss or triplet loss, depending on whether the goal is categorical identification or continuous similarity measurement. This localized approach improves overall reliability while keeping training complexity manageable by avoiding uniform application of complex loss functions across all parts.
Solution Approach 2:
The training process employs multiple loss functions with different parameter configurations optimized for specific parts. By adjusting loss function parameters and weights based on the importance and characteristics of each facial part, the system achieves higher reliability in matching critical features while balancing the overall training complexity through parameter optimization rather than structural complexity.
Data Source
AI summary
Systems, and method and computer readable media that store instructions for high-resolution face recognition.