基于投影梯度的NMF相关反馈医学图像检索算法

摘要

Relevance feedback technique has been an important approach in image retrieval. Because of including human's participation, it can reparation the disadvantage that the basic features of image cannot represent its semantic content well.The algorithm constructs a Non-negative Matrix Factorization basic matrix, which can be used to find more relevant images in the whole image database. A novel relevance feedback algorithm is presented based on projected gradient methods for NMF. Compared with the popular multiplicative update approach for NMF, in the assurance of Precision ratio and Recall ratio, it can improve the speed of the retrieval. Experiments were carried out on a big size database of 586 cerebral hemorrhage images, it shows the feasibility of this method.%相关反馈技术是近年来在图像检索中较为重要的研究方法,由于有人的参与,它能在一定程度上弥补图像的底层特征难以表达图像语义内容的不足.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,因而对整个图像库进行检索时可以查找到更多的相关图像.提出了一种基于投影梯度的非负矩阵分解(NMF)相关反馈方法,与常用的基于乘法更新的NMF相比,在保证查准率、查全率基本不变的情况下,能大大的提高反馈速度.使用由586幅脑出血图像组成的图像库进行实验,结果表明该方法的可行性.