Image Annotation Method Based on Graph Volume Network

The emergence of image annotation technology solves the problem of the rapid search of massive image data. This paper proposes a multi-label image annotation method based on graph convolutional network (GCN). First, it constructs a directed graph on the predictive object label to capture the dependencies between the labels and a set of interdependent object classifiers. The Resnext network is then used to extract image features, and the classifier is applied to the classification of extracted features, making the entire network end-to-end trainable. Finally, experiments on the Corel 5K and IAPR TC12 datasets show that the correct rate of image annotation is increased by 6.7% and the system execution speed is increased by 5.75%.