Bilinear Neighborhood Discriminant Embedding and Its Non-Iterative Solution Algorithm

This paper explores the use of Bilinear Neighborhood Discriminant Embedding (BLNDE) as a means to improve the performance of face recognition. BLNDE, which is based on graph embedding, utilizes the original image data directly and takes into account the row and column correlations of image matrix. By constructing the local within-class and between-class nearest-neighbor graphs and introducing the uncertainty factor, which is proposed as the individual property to describe the contribution to classification of each sample, into the process of dimensionality reduction, BLNDE achieves the purpose to gather the within-class samples and separate the between-class samples in feature subspace. Experiments are performed using the ORL face data set, which is proposed by Olivetti Research Laboratory. Results indicate that BLNDE outperforms Generalized Low Rank Approximations of Matrices (GLRAM) and Two-Dimensional Linear Discriminant Analysis (2DLDA).