Image Annotation Based on Semantic Structure and Graph Learning
Zhikui Chen,Meng Wang,Jing Gao,Peng Li
DOI: https://doi.org/10.1109/dasc-picom-cbdcom-cyberscitech49142.2020.00085
2020-01-01
Abstract:Image annotation is an important method to mine semantic information of images. The current methods do not fully consider the semantic repetitiveness and the imbalance hidden in labels, resulting in the unsatisfied image annotation. To address those problems, a semantic-independent nearest-neighbor graph model is proposed based on semantic structure and graph learning. Specifically, graph learning is used for producing the pre-annotation of images on the basis of label propagation of nearest-neighbor images, which can improve accuracy of weak labels. Then, the semantic structure and the word graph are introduced to fine-tune the image annotation, which can reduce the redundancy of the predicted labels. Finally, two-representative datasets are used to evaluate the proposed method. The results show that the proposed method outperforms the compared methods in terms of precision, recall, F1 and N+.