基于流形学习的图像检索算法研究

流形学习以发现非线性高维数据的本质维数为目标,使其更适合数据分析和高维数据的降维。图像检索中“语义鸿沟”问题指的是高维数据空间与低维的语义子空间之间的鸿沟,虽然利用相关反馈机制可以缩小这种鸿沟提高准确率,但是因为反馈图像数目较少,图像特征维数相对较高,会容易产生维数灾难问题。流形学习的引入为解决这一难题带来了新的希望,因为通过流形学习的方法学习高维图像特征数据的本征维数用于图像检索,大大提高了检索性能。基于流形学习的图像检索算法都是半监督的流形学习,充分利用了反馈信息,学习查询图像的语义子空间,有效的实现了高维数据的降维。

Image retrieval algorithms based on manifold learning

The purpose of the manifold learning is to discover the intrinsic dimensions of nonlinear high-dimensional data, which makes it more suitable for data analysis and dimensional reduction. The gap between high-dimensional data space and low-dimensional semantic subspace forms the “semantic gap” problem in image retrieval. Although using relevance feedback mechanism can narrow down the gap and increase the retrieval accuracy, the limitations of relevance feedback and the high dimensionality of image features make it prone to the course of dimensionality. Manifold learning has brought promise for settling these problems. Using the learned intrinsic dimensions of high dimensional image feature data by manifold learning can considerably enhance retrieval performance. The image retrieval algorithms based on manifold learning all take semi supervised learning strategy. It makes the most of the feedback information to learn the semantic subspace of image, and reduces the high dimensionality effectively.