基于非负矩阵分解的相关反馈图像检索算法

Non-negative Matrix Factorization Based Relevance Feedback Algorithm in Image Retrieval

提出了一种新的基于非负矩阵分解(NMF)的相关反馈检索算法.在每次反馈过程中,由用户标记与查询图像相似的正例样本的特征向量构成样本矩阵,进行NMF分解,得到NMF的基矩阵和样本的系数矩阵,然后根据分解所得的模型进行检索.由于NMF在一定程度上勾勒出了相关图像在基矩阵所代表的空间中的分布,因而可以有效地提高检索的查准率.使用由500幅图像组成的图像库进行实验,通过与特征加权以及支撑向量机相关反馈方法的比较表明,该方法通过交互的NMF相关反馈,确实能使图像检索的查准率得到较大的提高.

A novel relevance feedback algorithm was presented based on non-negative matrix factorization (NMF) learning in content-based image retrieval system. During the retrieval process, users can mark images similar to the query image as positive samples. Then the algorithm constructs an NMF basic matrix with the eigen vectors of the positive samples, which can be used to increase the accurate ratio for the image retrieval. Experiments were carried out on a big size database consisting of 500 images. The results show that accurate ratio of image retrieval can be increased much after using interactive NMF feedback algorithm.