| Automatic image annotation has increasingly exerted a tremendous fascination on researchers with the development of Internet and multimedia information technology.Since the image annotation methods based on various graph structures were proposed,image annotation performance has been dramatically improved.However,the existing image annotation algorithms based on graph structure are divided into two categories: the model with graph regularization items and the model with graph convolutional network(GCN).These two types of algorithms do not make full use of the relationships among data.It loses sight of semantic information when constructing the sample graph(the relationship among samples).Similarly,it ignores visual content when constructing the label graph.Either one only includes a sample graph or just a label graph.Therefore,it is formidable to accurately describe the data’s geometry,which affects the labeling effect.Meanwhile,it induces complicated calculation and time-consuming when constructing multiple graph structures.The paper promotes the automatic image annotation algorithm’s performance based on graph structure from building a more accurate graph structure,improving annotation performance,and reducing computational complexity.The main work is as follows:1.An improved image annotation algorithm using Laplacian graph structure embedding is proposed.Laplacian eigenmaps are adopted to construct a sample graph in the case of missing tags.It uses all samples in the training set to group a Laplacian matrix and embeds the Laplacian matrix into the feature mapping.Then,the original manifold information will maintain after dimensionality reduction.Experimental results conducted on the two image annotation datasets demonstrate the effectiveness of the proposed method.2.The two improved annotation methods based on graph structure embedding combined with semantic information is proposed and both of which consider semantic information when constructing sample graphs.Therefore,the sample graph is more accurate.Meanwhile,the two annotation methods respectively employ the probability Laplacian graph constructed by prior knowledge and the Laplacian graph obtained by the relationship between the cluster centers to represent tag correlations.Experimental results show that the two annotation methods achieve superior performance on the benchmark image annotation datasets to state-of-art methods.3.An image annotation method based on central attention with multi-graph embedding is proposed.We assume that the center part of an image is more significant.The graph structure formed by the center parts can express the relationship among samples and pay more attention to the detailed information.Therefore,it is accurate and reduces the computational complexity relying solely on the two graph structures composed of the central and global features.Experimental results show that this method achieves promising annotation performance and saves program running time.4.An image annotation method using a parallel graph convolutional neural network is proposed.This method combines a convolutional graph network with image annotation to improve the annotation performance under semi-supervised learning.We also introduce the parallel graph convolutional neural network that connects the label graph and the sample graph.The two graph structures complement each other to improve the annotation results further.The experimental results prove the excellent performance of this method. |
