Learn to Rectify Label Through Kernel Extreme Learning Machine

Abstract

Recent studies attempt to construct complicated and redundant Convolutional Neural Networks (CNNs) to improve image classification performance. In this paper, instead of painstakingly designing a CNN’s architecture, we consider promoting classification performance by revising CNN’s classification results. We therefore propose a novel image classification approach that Learns to Rectify Label (LRL) through Kernel Extreme Learning Machine (KELM). It includes two phases: (1) Pre classification, we put images into a trained CNN to generate corresponding incomplete labels. (2) Label Rectification, the incomplete labels are rectified by the KELM’s high-dimensional mapping, so final classification results are acquired. Extensive experiments conducted on public datasets demonstrate the effectiveness of our method. At the meantime, our method has well generalizability that can be integrated with many popular networks.