Abstract: The image annotation method based on the probabilistic topic model annotates images by learning the semantic of the image, and researchers pay more and more attention to it in recent years. Class label information can provide the valuable information for image annotation, for example, for images in “tall building” class, annotating “sky”, “skyscraper” is more possible than annotating “sea” and “beach”. However, for images in “coast” class, annotating “sea”, “beach” is more possible than annotating “sky” and “skyscraper”. This paper proposes an image annotation probabilistic topic model fusing class information which uses class information to promote image annotation. And it derives a parameters estimation algorithm based on the variational EM algorithm, as well as gives the method annotating the new images. The experimental results on LableMe and UIUC-Sport datasets show that the image annotation performance of the proposed model is better than other contrastive models.

Key words: image annotation, image class, variational expectation maximization, Corr-LDA model

摘要： 基于概率主题模型的图像标注方法旨在通过学习图像语义进行图像标注，近年来倍受研究人员关注。考虑到类别对图像标注可提供有价值的信息，例如，“高楼”类图像，出现“天空”、“摩天楼”的可能性大于“海水”和“沙滩”。而“海岸”类图像出现“海水”、“沙滩”的可能性要大于“天空”和“摩天楼”。在Corr-LDA模型的基础上利用图像类别来改进图像的标注性能，提出了一个融入类别信息的图像标注概率主题模型。为该模型推导了一个基于变分EM的参数估计算法，并给出了使用该模型标注图像的方法。在LabelMe和UIUC-Sport两个真实数据集上验证了提出模型的标注性能要高于其他相比较模型。

关键词: 图像标注, 图像类别, 变分EM, Corr-LDA模型