Graph steered discriminative projections based on collaborative representation for Image recognition

摘要

Dimensionality reduction techniques are commonly used for image recognition. We propose a graph steered dimensionality reduction method called Discriminative Projections based on Collaborative Representation (DPCR) by transforming the dimensionality reduction task into a graph embedding framework. DPCR utilizes the collaborative representation to construct within-class and between-class graphs. To improve the discriminative performance of dimensionality reduction, DPCR introduces the label information into graph building. The novel method not only avoids the difficulty of finding proper neighborhood but also inherits the merits of manifold learning methods and the robustness of collaborative representation techniques. Experiments on benchmark datasets demonstrate its effectiveness.