Abstract
This paper introduces a novel framework for face recognition that combines a "Global-Local" (Glocal) tensor representation of images with a new discriminative orthogonal rank-one tensor decomposition (DORTD) method. By preserving spatial structures and enforcing orthogonality between projections, the approach achieves state-of-the-art accuracy with significantly more compact embeddings than traditional subspace methods. We evaluate the proposed method on benchmark face databases including Yale, ORL, and CMU PIE, demonstrating its superior generalization and physical interpretability.
Introduction
Traditional face recognition algorithms, such as Eigenfaces (PCA) or Fisherfaces (LDA), typically "flatten" 2D images into 1D vectors. This process destroys the inherent spatial relationships between pixels and often leads to the "curse of dimensionality," where the number of parameters to estimate far exceeds the available training data. While previous tensor-based methods attempted to solve this by treating images as multi-dimensional arrays, they often lacked physical interpretability or suffered from overfitting. This paper addresses two specific gaps: the loss of spatial structure during representation and the need for better generalization via orthogonality constraints.
Global-Local (Glocal) Tensor Representation
A primary innovation of this work is the Glocal tensor representation. Instead of using the raw 2D image, the image is partitioned into non-overlapping blocks. These blocks are raster-scanned into vectors and placed as columns in a new 2D tensor.
In this representation, the left projection acts as a set of local filters applied across the image, while the right projection determines which spatial blocks (e.g., eyes vs. forehead) are most important for recognition. This ensures that the learned projections correspond to meaningful physical features, unlike standard PCA where "eigenfaces" can appear as uninterpretable blobs.
Discriminative Orthogonal Rank-one Tensor Decomposition (DORTD)
The paper presents a method to find a set of rank-one tensor projections that are both discriminative (maximizing between-class variance while minimizing within-class variance) and orthogonal to one another.
The author proves a theorem showing that two rank-one tensor projections are orthogonal if their constituent vectors are orthogonal in at least one dimension. Since there is no closed-form solution, a sequential iterative optimization algorithm is used to solve a generalized eigenvalue problem at each step. Enforcing orthogonality acts as a powerful regularizer, reducing the model's capacity (VC-dimension) to prevent it from "memorizing" noise in the training set, thereby leading to better generalization on unseen data.
Experimental Results
The method was evaluated on four classic face recognition benchmarks: Yale, ORL, Extended YaleB, and CMU PIE. The proposed ORO$_{4\times2}$ (Orthogonal Rank-one with 4x2 block partitioning) consistently outperformed or matched state-of-the-art methods:
| Dataset | Method | Accuracy | Dimensions |
|---|---|---|---|
| Yale | ORO$_{4\times2}$ | 86.8% | 94 |
| ORL | ORO$_{4\times2}$ | 97.0% | 105 |
| YaleB | ORO$_{4\times2}$ | 91.0% | 108 |
| PIE | ORO$_{4\times2}$ | 93.6% | 73 |
Notably, on the Yale dataset, ORO$_{4\times2}$ achieved 86.8% accuracy, significantly higher than standard LDA (77.5%) or LPP (78.3%). On the PIE dataset, it matched the accuracy of the Orthogonal Locality Preserving Projections (OLPP) method but used only 73 dimensions instead of 381, highlighting its extreme compactness.
Discussion and Conclusion
The DORTD framework offers a robust and interpretable approach to face recognition. By leveraging the Glocal representation and enforcing orthogonality, the method achieves high accuracy with very few dimensions. However, the author notes that for extremely small training sets, the orthogonal regularization might be too strong, potentially limiting the model's ability to learn complex patterns. Additionally, the iterative optimization process is more computationally intensive than simple linear projections, though it remains efficient for large-scale databases once trained. Overall, the resulting model generalizes better and provides concrete physical insights into the features used for recognition.
Summary updated