Unsupervised feature selection using orthogonal locality preserving projections and bipartite graph matching for face image classification

Feature selection plays a crucial role in facial image classification by reducing dimensionality and improving robustness to variations in expression, pose, and lighting. However, researchers face challenges when selecting features from high-dimensional, unlabeled data due to the nonlinear manifold structure of facial images. To address this, this paper proposes UFSOLPP, a novel unsupervised feature selection method that consists of three main stages. First, the method employs Orthogonal Locality Preserving Projections (OLPP) for feature extraction, aiming to preserve local data structures and enforce orthogonality without dimensionality reduction. Unlike conventional OLPP, which uses heat kernel to measure similarity, this paper replaces it with cosine distance to better capture angular relationships that are for facial image discrimination. Second, it measures the similarity between the original and orthogonal features using the Pearson correlation distance. Third, it models both feature sets as vertices in a weighted bipartite graph. The edge weights are computed using the Pearson correlation similarity, and the method uses the Hungarian algorithm to compute maximum matching. The method selects the original features involved in the maximum matching as the final subset. This strategy removes noisy, correlated, and redundant features effectively, while preserving interpretability and discriminative power. Experiments demonstrate that UFSOLPP outperforms eight state-of-the-art methods. It achieves ۹۶.۰۰% accuracy and ۰.۹۸۰۰ NMI on Jaffe, ۶۸.۶۶% accuracy and ۰.۷۵۳۲ NMI on ORL, and ۸۲.۳۳% accuracy and ۰.۸۵۵۷ NMI on the high-dimensional Pixraw۱۰P dataset. These results highlight the practical value of UFSOLPP and its ability to handle high-dimensional data efficiently in unsupervised facial image analysis.

کلیدواژه ها:

OLPP ، Manifold learning ، Weighted Bipartite Graph Matching ، Hungarian algorithm