基于超图直推非负矩阵分解的图像标注法研究

The Research of Image Annotation based on Hypergraph Transduction Non-negative Matrix Factorization

为解决传统图像标注方法难以准确建立从低层视觉特征到高层语义空间映射关系的问题,提出了一种基于超图直推非负矩阵分解的图像标注算法。通过把有监督超图正则化思想引入到非负矩阵分解框架,使得图像标注算法可以有效地利用样本间复杂的多元关系和标注信息,而直推学习正则项的利用又增加了算法对标签预测误差进行合理控制的能力。在图像标注数据集上的仿真结果表明,相对于支持向量机、鉴别式度量学习等传统的图像标注算法,提出的算法大幅提高了标注的准确率和模型的鲁棒性。并具有很好的可行性和有效性。

This article proposes an image annotation algorithm based on non - negative matrix factorization of hypergraph transduction. Firstly, the regularization principle of supervising hypergraph was introduced into frame of non - negative matrix factorization. The image annotation utilized complex multi - relationship among samples and annotation information effectively. Meanwhile, rational control ability of algorithm for forecast error of label was increased to use regular terms of transduction learning. Results of simulation on dataset of image annotation show that the algorithm improves accuracy of annotation and robustness of model significantly compared with traditional image annotation algorithms such as support vector machine and measurement learning with discrimination type. It has excellent feasibility and validity.