基于Nyström 低阶近似的半监督流形排序图像检索

实际图像检索过程中,用户提供的相关反馈有限,但存在大量未标记图像数据. 本文在前期半监督流形图像检索工作的基础上,提出一种基于Nystrm低阶 近似的半监督流形排序图像检索方法.通过采用半监督的流形正则化框架, 将图像数据嵌入到低维流形结构中进行分类排序,以充分利用大量未标记数据, 并兼顾分类误差、数据分布的几何结构以及分类函数的复杂性.针对半监督学习速度缓慢的问题, 基于Nystrm低阶近似对学习过程进行加速.在较大规模的Corel图像数据集上进行了检索实验, 实验结果表明该方法能获得较好的效果.

Semi-supervised Manifold-ranking-based Image Retrieval with Low-rank Nyström Approximation

In the real image retrieval process, there are abundant unlabeled images whereas there only exist few labeled images. To address this issue, based on our previous work of semi-supervised manifold image retrieval, this paper proposed a novel learning method named semi-supervised manifold ranking based image retrieval (S²MRBIR). The images are assumed always embedded in low-dimensional sub-manifolds. In particular, S²MRBIR adopts the manifold regularization framework to rank the retrieved images while regarding the relevant feedback process of image retrieval as an online learning process and treating the image retrieval as a classification problem. The manifold regularization framework is capable of taking account of both labeled and unlabeled data, the classification performance, the geometric structures of the data distribution, and the complexity of the classifier. Moreover, an accelerating algorithm based on Low-rank Nyström approximation was proposed to improve the computing procedure of S²MRBIR (NA-S²MRBIR). Experimental results on Corel image database demonstrated the effectiveness of S²MRBIR.