基于距离测度学习的AP聚类图像标注

发布时间：2018-03-13 10:53

本文选题：距离测度学习　切入点：近邻传播(AP)聚类算法　出处：《计算机工程与应用》2017年23期 　论文类型：期刊论文

【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。
[Abstract]:In this paper, an improved AP(Affinity Propagation clustering model is proposed to effectively utilize the image underlying visual features and image semantic features. Firstly, a semi-supervised distance measure learning algorithm is used to fuse the image semantic information. Then we use the new distance measure to cluster each kind of image by AP clustering, generate the clustering center of each kind of image, calculate the average distance between the image to be tagged and the cluster center of each kind of image. Finally, the distance between the image to be tagged and each cluster center within the cluster is calculated, and the category of the image to be labeled is determined, and the tagging words of the image under the category are counted. As the tagging words of the image to be tagged, the experiments are carried out on the Corel5K and NUS-WIDE data sets, and it is proved that the method can effectively improve the accuracy of the tagging.
【作者单位】：北京信息科技大学网络文化与数字传播北京市重点实验室;首都医科大学宣武医院;
【基金】：网络文化与数字传播北京市重点实验室开放课题(No.ICDD201504) 国家自然科学基金(No.61271304) 北京成像技术高精尖创新中心项目(No.BAICIT-2016003) 2014年度国家社会科学基金委托课题(No.14@ZH036)
【分类号】：TP391.41

【相似文献】