基于对偶主题空间的PLS图像标注方法

PLS Image Annotation Based on Dual Mode Semantic Space

为提高图像标注精度,针对如何融合图像文本特征和视觉特征的问题,提出基于对偶主题空间的图像标注技术。首先将图像的视觉特征与文本标注表示为同一对象的两种视图方式,运用偏最小二乘(PLS)的多元统计分析理论,考虑两个特征空间之间的语义对偶关系,抽取得到双模态共有语义信息;然后在双模主题构成的对称空间上构建一个非概率主题标注模型;最后,待标注图像根据视觉特征在对偶主题空间的投影计算出标注的预测向量,设置阈值得到标注关键词。在公共数据集Core5K上进行算法性能的测试,实验表明,基于对偶主题空间的标注算法可以有效提高图像标注的性能和标签准确的个数。

In order to improve the accuracy of image annotation, and solve the problem of how to integrate text features and visual features of images, this paper puts forward an image annotation algorithm based on dual theme space. Firstly, the visual features and text annotation of image are represented as two views of the same object. Based on the multiple statistical analysis theory of partial least squares(PLS), the semantic dual relationship between the two feature spaces is considered to extract the dual shared semantic information. Then, a non-probabilistic annotation model is constructed on the symmetric space composed of dual theme. Finally, the predicted annotation vectors are calculated by the projection of visual features on dual theme space, the annotation texts of a new image are selected by setting the threshold. The algorithm performance is tested on the public data set of Core5 K, experiments show that the proposed algorithm based on dual mode semantic space can effectively improve the performance of image annotation and the number of annotation accuracy.