2D LDA Face Recognition Spatial Matrix Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current face recognition methods, such as PCA and LDA, face challenges in efficiently distinguishing between class variations and achieving robustness against facial expressions and pose changes, while also dealing with high computational and memory requirements.
Innovation Solution
The proposed method, Bidimensional Linear Discriminant Analysis (LDA2D), hierarchizes a learning base into classes, forming matrices to represent variance between and within classes, and determines eigenvectors to minimize intra-class variance and maximize inter-class variance, reducing dimensionality and computational load.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional LDA is used to maximize discrimination between classes, then recognition accuracy is improved, but computational complexity and memory requirements increase
Solution Approach 1:
The patent transitions from traditional 1D LDA to 2D LDA by preserving the spatial structure of face images. Instead of flattening images into vectors, the method operates directly on 2D matrices, performing discriminant analysis in the spatial domain. This dimensional change reduces computational complexity while maintaining discrimination accuracy by exploiting the inherent 2D structure of facial images.
Solution Approach 2:
The patent segments the face image data into structured 2D blocks rather than treating them as flat vectors. By maintaining the spatial segmentation of rows and columns in the image matrix, the method enables more efficient computation through block matrix operations, reducing both memory requirements and computational load while preserving class discrimination capabilities.
2Productivity
If PCA is used for dimension reduction, then computational load is reduced, but discrimination between classes deteriorates
Solution Approach 1:
The patent applies 2D LDA that operates in the spatial domain rather than transforming to 1D eigenvector space as in PCA. By performing discriminant analysis directly on 2D image matrices and preserving spatial relationships, the method achieves both computational efficiency and superior class discrimination compared to PCA, which loses spatial structure during vectorization.
3Measurement precision
If more components are retained in PCA/LDA to improve recognition accuracy, then discrimination is improved, but memory requirements increase
Solution Approach 1:
The patent reduces memory requirements by operating in the 2D spatial domain rather than requiring storage of numerous 1D eigenvector components. The 2D LDA approach compresses face recognition information into a compact set of 2D projection matrices that preserve spatial structure, significantly reducing the memory space needed while maintaining or improving recognition accuracy.
Data Source
AI summary
The invention relates to a method for recognizing faces in digital images consisting in providing for a knowledge base which contains face images and is hierarchical into several classes each of which comprises different images of the same person. The invention relates to pre-processing said knowledge base in such a way that a minimization of variance in each class and a maximization of variance between different classes are simultaneously obtainable, thereby making it possible to form a vectorial base comprising the discriminant component of said knowledge base. The comparison of a recognizable face with a pre-processed reference face such as the knowledge base and an eventual reconstruction of a recognized face are also disclosed.