Two-Dimensional PCA Face Recognition Source Code

Two-dimensional Principal Component Analysis (2D-PCA) serves as an efficient feature extraction method for face recognition applications, particularly well-suited for image data processing. When applied to the ORL face database (containing 10 grayscale images per subject with varying poses/illumination for 40 individuals), this approach achieves a remarkable 98% recognition rate through the following key steps:

Data Preprocessing ORL images are uniformly rescaled and converted into column vectors to form the original data matrix. Unlike traditional PCA that requires vectorization, 2D-PCA directly processes image matrices, thereby preserving more spatial structural information.

Covariance Matrix Computation The covariance matrix is calculated along the column direction of image matrices, avoiding the computational burden caused by high-dimensional vectors in conventional PCA. Feature values and eigenvectors are extracted through Singular Value Decomposition (SVD).

Projection and Dimensionality Reduction Select the eigenvectors corresponding to the top k largest eigenvalues to form a projection matrix. The original images are then projected onto a low-dimensional subspace, significantly reducing data dimensionality while retaining crucial discriminative features.

Classification and Recognition A nearest neighbor classifier (e.g., using Euclidean distance) matches projected features between test samples and training samples. When validated through cross-validation methods (such as leave-one-out) on the ORL database, the 98% recognition rate demonstrates 2D-PCA's robustness against illumination and pose variations.

Advantages and Extensions Computational Efficiency: 2D-PCA directly processes image matrices, reducing computational complexity by approximately one order of magnitude compared to traditional PCA. Interpretability: The projected eigenvectors can be reversely visualized as "eigenfaces," intuitively reflecting key discriminative regions.

This method can be efficiently implemented in MATLAB using matrix operation functions (e.g., `svd` for singular value decomposition), making it suitable as an introductory practice solution for image feature dimensionality reduction.