基于语义链接网的图像自动标注

Image Auto-Annotation Using Semantic Link Network

摘要: 自动图像标注是机器视觉领域中的一项具有挑战性的课题。大多数传统方法聚集在视觉图像与标注概念间的建模，而忽略了语义概念之间的关系。本文提出了一种新的标注方法IA-SLN，它能有效捕获语义概念关系从而提升图像标注性能。首先，构建了一个基于全局关联语义与局部关联语义学习的语义链接网。其次，利用它标注未标注图像。当一个标注概念与其关联概念们频繁共现于图像中，IA-SLN方法将提升该标注概念的语义预测值。最后，通过一个标注增强新策略获得未标注图像的最相关语义标注。在IAPR公共数据集上对比了其他方法，实验结果表明我们的图像标注方法IA-SLN性能更优。

Abstract: Image Auto-annotation remains as a challenge in machine vision. Most conventional approaches concentrate on the relations between visual images and labeled concepts, and neglect the correla-tions between semantic concepts. In this paper, we present a novel approach called Image Au-to-annotation using Semantic Link Network (IA-SLN), which can effectively capture semantic cor-relations to boost image auto-annotation. Specifically, we first construct a semantic link network based on the global concept correlations and the local concept correlations. Second, we utilize it to annotate unlabeled images. When a concept and its associated concepts frequently occur in images, the prediction of this concept will be boosted in our IA-SLN. Finally, we obtain the most relevant annotations for each of unlabeled images by using a new strategy of annotation promotion. The experimental results on the publicly available dataset IAPR have shown that our approach performs favourably compared with several other state-of-the-art methods.