Enhanced Local Receptive Fields based Extreme Learning Machine using Dominant Patterns Selection

The local receptive fields based ELM (ELM-LRF) is an extended version of ELM. Its hidden nodes are structure through the local connection approach, which demonstrated satisfactory performance in image classification problems. However, ELM-LRF still requires further improvement because extracting images features directly with random initial weights will generate redundancy features that may degrade its performance in some situations. This paper, therefore, presents a new method named the enhanced local receptive fields based ELM using dominant patterns selection (DP-ELM-LRF) to enhance ELM-LRF, which applies novel feature selection in vehicle detection through the selection of dominant patterns of HOGs (DPHOG) for selecting dominant features in the ELM feature space. DP-ELM-LRF evaluated classification performance on GTI and Concrete Crack datasets for binary classification and MNIST, Semeion, and small NORB datasets for multi-classification. Experiment results demonstrated that the DP-ELM-LRF was superior to the ELM-LRF and other comparative methods of multi-classification, whereas binary classification, DP-ELM-LRF, remains comparable with ELM-LRF.