基于协作式标注图像数据的垃圾标签检测方法

Spam Tag Detection Method Based on Collaborative Annotation Image Data

由于用户标签的不准确和语义模糊使得协作式标注图像检索正确率低,而现有垃圾标签过滤方法往往关注标签本身,忽略了协作式标签与图像的关联性。本文在分析协作式标注图像视觉内容与标签的关联性的基础上,提出一种基于协作式标注图像视觉内容的垃圾标签检测方法。该方法分析同一标签下图像视觉内容,设计不同的核函数用于颜色和SIFT(Scale-invariant feature transform)特征子集,同时将2种低维特征映射到高维多模特征空间形成混合核函数,对同一标签下的图像进行基于混合核的最大最小距离聚类,少数群体的标签说明与图像内容关联性小则为用户标注错误的标签,从而检测垃圾标签。实验结果表明,该方法能够提高协作式图像垃圾标签检测的正确性。

The accuracy of the collaborative tagging image retrieval is lower because of the inaccuracy of user’ s annotation. Exist-ing spam tag detection methods tend to focus on label itself, ignoring the correlation between collaborative label and image. Ana-lyzing the correlation of collaborative tagging image visual content and image tags, the spam tag detection method of collaboration annotation based on visual content of collaborative tagging image is proposed. The method analyze visual content of images which have the same tag and design different kernel functions for color and SIFT feature subset. The two features will be mapped form low dimensional space to high dimensional character space, while the mixed-kernel function is established. Finally, the images which have the same tag is clustered by max-min distance means, and the tag of images in the class which has a few images are spam tags because of weak correlation. The experimental results show that the method can improve the accuracy of the tag spam detection on collaborative annotation images.