Face Recognition via Multi-Model Error Decomposition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face recognition systems face challenges in accurately identifying subjects under varying conditions such as occlusions, illumination, and expression variations, as they often rely on assumptions about noise distribution that may not hold in all cases, leading to reduced recognition rates when faced with complex occlusions and mixed noise types.
Innovation Solution
A face recognition method that represents test samples as a combination of training face images and error terms, estimating the coefficient vector and error terms through a constrained optimization problem, and computes residual errors to classify the face sample by selecting the face class with minimal residual error, using a robust error correction multi-model representation approach.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a single error model assumption is used for noise correction, then the method is simple and computationally efficient, but it fails to accurately characterize complex occlusions and mixed noise types leading to reduced recognition rates
Solution Approach 1:
The patent divides the error characterization into multiple independent components: low-rank error term for structured occlusions, sparse error term for random corruptions, and dense error term for illumination variations. Each error type is modeled separately with its own regularization term, allowing the system to accurately characterize complex noise patterns while maintaining computational tractability through modular optimization.
Solution Approach 2:
The patent combines multiple error modeling approaches into a unified composite framework. The total error is decomposed into low-rank + sparse + dense components, each handled with appropriate regularization. This composite error model integrates the strengths of different approaches (low-rank for block occlusions, sparse for random noise, dense for lighting) to create a more robust and accurate characterization of real-world face image corruptions.
2Reliability
If low-rank regularization is applied to characterize occlusions, then block-occlusion recognition improves, but recognition rates drop dramatically when corrupted pixel percentage increases significantly violating the low-rank assumption
Solution Approach 1:
The patent segments the error characterization by introducing multiple error components with different regularization properties. The low-rank term handles structured block occlusions, while the sparse term handles random pixel corruptions. This segmentation allows the model to adapt to different corruption patterns and maintain performance across a wide range of corruption levels without violating underlying assumptions.
Solution Approach 2:
The patent changes the regularization parameters dynamically based on the corruption pattern. For low-rank errors (block occlusions), nuclear norm regularization is applied. For sparse errors (random corruptions), l1-norm regularization is used. This parameter adaptation allows the model to maintain effectiveness across varying corruption levels and types.
3Measurement precision
If sparse representation with optimization methods is used, then face recognition accuracy improves, but computational complexity increases compared to simpler classification methods
Solution Approach 1:
The patent changes the regularization norm from l1 (sparse) to l2 (dense) for the representation coefficients. This parameter change simplifies the optimization problem significantly, allowing closed-form solution without iterative optimization methods. The l2-regularized least squares formulation maintains good recognition accuracy while dramatically reducing computational complexity compared to l1-sparse representation.
Solution Approach 2:
The patent replaces the iterative optimization mechanism (required for l1-sparse representation) with a direct closed-form mathematical solution (using l2-regularization and normal equations). This substitution eliminates the need for complex iterative algorithms while maintaining recognition accuracy, significantly simplifying the computational process.
Data Source
AI summary
The present invention provides a face recognition method on a computing device, comprising: storing a plurality of training face images, each training face image corresponding to a face class; obtaining one or more face test samples; applying a representation model to represent the face test sample as a combination of the training face images and error terms, wherein a coefficient vector is corresponded to the training face images; estimating the coefficient vector and the error terms by solving a constrained optimization problem; computing a residual error for each face class, the residual error for a face class being an error between the face test sample and the face test sample's representation model represented by the training samples in the face class; classifying the face test sample by selecting the face class that yields the minimal residual error; and presenting the face class of the face test sample.