Graph-preserving shortest feature line segment for dimensionality reduction

Shortest feature line segment (SFLS) is a recently proposed classification approach based on nearest feature line (NFL). It naturally inherits the representational capacity enlargement property of NFL and offers many other benefits in accuracy and efficiency. However, SFLS still has several drawbacks, limiting its generalization ability. In this paper, we develop a manifold learning algorithm for dimensionality reduction based on a novel line-based metric derived by integrating SFLS and NFL, which takes advantage of the benefits of the two algorithms and avoids their disadvantages. Unlike the construction of a point-based relationship in traditional dimensionality reduction algorithms, the new measurement forms linear models of multiple feature points, which capture more information than individual prototype and serve to discover the intrinsic connection of nearby points. Moreover, to enhance the discriminating capability, the affinity matrix in graph embedding is designed in supervised manner by using class label information. Experimental results on four standard databases for face recognition confirm the effectiveness of our proposed method.