基于图嵌入的自适应多视降维方法

随着监控摄像头的普及和数据采集技术的快速发展，多视数据呈现出规模大、维度高和多源异构的特点，使得数据存储空间大、传输慢、算法复杂度高，造成“有数据、难利用”的困境。到目前为止，国内外在多视降维方面的研究还比较少。针对这一问题，本文提出一种基于图嵌入的自适应多视降维方法。该方法在考虑视角内降维后数据重构原始高维数据的基础上，提出自适应学习相似矩阵来探索不同视角之间降维后数据的关联关系，学习各视数据的正交投影矩阵实现多视降维任务。本文在多个数据集上对降维后的多视数据进行了聚类/识别实验验证，实验结果表明基于图嵌入的自适应多视降维方法优于其他降维方法。

An adaptive multi-view dimensionality reduction method based on graph embedding

With the popularity of surveillance cameras and the rapid development of data acquisition technology, multi-view data shows the traits of large scale, high dimension and multi-source heterogeneity, which cause large data storage, low data transmission speed and high algorithm complexity, resulting in a predicament that “there are plenty of data that are hard to use”. Up to now, few domestic and foreign researches have been done on multi-view dimensionality reduction. In order to solve this problem, this paper proposes an adaptive multi-view dimensionality reduction method based on graph embedding. In consideration of the reconstructed high-dimensional data after the view-angle dimensionality reduction, this method puts forward an adaptive similarity matrix to explore the correlation between dimension-reduced data from different perspectives and learn the orthogonal projection matrix of each perspective to achieve the multi-view dimensionality reduction task. In this paper, a clustering/recognition verification experiment is performed on the dimension-reduced multi-view data from multiple data sets. The experimental results present that the proposed method is better than other dimensionality reduction methods.