Abstract: It annotates the image for effective use of image low-level visual feature and semantic feature and puts forward a kind of improved affinity propagation cluster annotation model. Firstly, it uses the semi-supervised distance metric learning algorithm to acquire new distance measure after training combined with image semantic information. Then it uses new distance measure to cluster image with affinity propagation for every kind of images, generates cluster center of various kinds of images, calculates the average distance from will-be annotated images to cluster center of various kinds of images and determines the category of will-be annotated images. Finally, it calculates the distance from will-be annotated images to various kinds of intra-class cluster center, determines the intra-class category of will-be annotated images and summarizes the marked words of this kind image as marked words of will-be annotated images. After experiments on databases of Corel5K and NUS-WIDE, it is verified that this method is successful to improve the accuracy of image annotation.

Key words: distance metric learning, Affinity Propagation（AP）, image annotation

摘要： 针对有效利用图像底层视觉特征和图像语义特征进行图像标注，提出一种改进的AP（Affinity Propagation）聚类标注模型。首先采用半监督距离测度学习算法，融合图像语义信息，训练得到新的距离测度。然后使用新的距离测度对每一类图像进行AP聚类，生成各类图像的聚类中心，计算待标注图像到各类图像聚类中心的平均距离，确定待标注图像类别。最后计算待标注图像到类内各个聚类中心的距离，确定待标注图像类内类别，统计该类别下图像的标注词，作为待标注图像的标注词。在Corel5K和NUS-WIDE数据集上进行了实验，经验证，该方法有效提高了标注精度。

关键词: 距离测度学习, 近邻传播（AP）聚类算法, 图像标注