Discriminative globality-locality preserving extreme learning machine for image classification

Abstract

Extreme learning machines (ELM) have been widely used in classification due to their simple theory and good generalization ability. However, there remains a major challenge: it is difficult for ELM algorithms to maintain the manifold structure and the discriminant information contained in the data. To address this issue, we propose a discriminant globality-locality preserving extreme learning machine (DGLELM) in this paper. In contrast to ELM, DGLELM not only considers the global discriminative structure of the dataset but also makes the best use of the local discriminative geometry information. DGLELM optimizes the projection direction of the ELM output weights by maximizing the inter-class dispersion and minimizing the intra-class dispersion for global and local data. Experiments on several widely used image databases validate the performance of DGLELM. The experimental results show that our approach achieves significant improvements over state-of-the-art ELM algorithms.