Multiple Graphs Structure of Intra-class Representation with Semantic Information for Image Annotation

Abstract

Automatic image annotation has absorbed a few research interests in recent decades quietly. Nevertheless, most methods leave graph structure out of consideration or utilize merely the graph without semantic information to capture manifold information of samples. Furthermore, it is inaccurate to consider the label matrix of a dataset as the feature matrix of tags during the computing label graph. To overcome these shortcomings, we propose a method, called Multiple Graphs Structure of Intra-class Representation with Semantic Information for Image Annotation. Our method proposes a superior multiple graphs structure called multiple graphs structure of intra-class representation as combining the semantic information with sample graph structure. Then, the probability Laplacian graph is presented to obtain the relationship between tags according to prior knowledge instead of exploiting tag features. Afterward, some experiments have been implemented to certify the effectiveness of the proposed graph structures on Corel5K, ESPGame, IAPRTC12, and NUS-WIDE datasets.