Central Attention with Multi-graphs for Image Annotation

Abstract

Automatic image annotation has exerted a tremendous fascination on researchers in recent decades due to the development of multimedia and computer vision. Much of the research only exercise the single graph to annotate the images in semi-supervised learning. And a multitude of methods examine how to use multi-view or image segmentation to compose multiple graph structures. However, the single graph is arduous to express the complete manifold structure information. Further, it is structurally computationally intensive and consumes a lot of time that building multiple graph structures depending on multiview or image segmentation. To this end, we propose a method, called Central Attention with Multi-graphs for Image Annotation. We firstly propose that the center part of an image is fatal for the image. And the promising performance can be accomplished relying solely on the two graph structures composed by central and overall features in semi-supervised learning. A number of experiments have been conducted on benchmark datasets, such as Corel5K, ESPGame, and IAPRTC12 to prove the effectiveness of the proposed method.