Fusion of CNN and ORB Detection with Graph Convolutional Network for Enhancing Feature Discrimination

Abstract

A novel image classification scheme called ORB-GCN, which combines convolutional neural networks (CNNs) with features from Oriented and Rotated BRIEF (ORB) detection for a graph convolutional network (GCN). CNN models often encounter challenges in differentiating between similar features, leading to reduced interpretability and lower accuracy. Enhancing feature discrimination in local and global information is the goal of the ORB algorithm fusion for graph construction in GCN. By training CNN and ORB-GCN simultaneously and performing end-to-end classification, the proposed method effectively improves the discriminative ability of features compared to state-of-the-art methods. According to experiments on the MIT Indoor CVPR09 and Intel Image Scene datasets, the proposed ORB-GCN approach has the best accuracy.