Journal of Frontiers of Computer Science and Technology ›› 2017, Vol. 11 ›› Issue (3): 414-426.DOI: 10.3778/j.issn.1673-9418.1603047

Previous Articles     Next Articles

Spectral Deep Learning Algorithm for Image Sequence

YIN Hongwei, LI Fanzhang+   

  1. College of Computer Science and Technology, Soochow University, Suzhou, Jiangsu 215000, China
  • Online:2017-03-01 Published:2017-03-09



  1. 苏州大学 计算机科学与技术学院,江苏 苏州 215000

Abstract: To better understand the hidden depth information of image sequence, people need to figure out the hidden data structure of the image sequence. At present, spectral manifold learning is efficient to learn the low dimensional embedding coordinates projected from the high dimensional sample data and thereby learning the manifold structure. Since the conventional method based on the hypothesis which the sample data are distributed on one single manifold, does not support hybrid multiple manifold model. Combined with the structure characteristics of image sequence, this paper presents a new algorithm called spectral deep learning (SDL) algorithm. Through setting up multimanifold mixed model, this algorithm preserves the continuity and smoothness of the local changes on the manifold. Through a method called manifold alignment, this algorithm establishes the mapping between the hierarchy manifold to find the deep low dimensional embedding coordinates of the image sequence. The experiments illustrate the validity of hybrid multiple manifold and image sequence data set of this algorithm.

Key words: image sequence, spectral manifold learning, hybrid multiple manifold, local tangent space, hierarchy  , manifold

摘要: 为了更好地理解图像序列的隐藏深度信息,需要分析数据的隐藏结构。目前,多采用谱流形学习算法学习高维采样数据的低维嵌入坐标,从而获取数据的隐藏结构。谱流形学习算法一般是基于所研究的高维数据分布在单个流形上的前提假设,并不支持图像序列中存在的多流形结构。结合图像序列的结构特点,提出了一种针对图像序列的谱深度学习算法(spectral deep learning,SDL)。通过建立混合多流形模型,保持流形局部变化的平滑和连续,利用流形对齐建立层次流形的映射关系,得到图像序列的深度低维嵌入坐标。最后通过实验证明了算法在混合多流形数据集和图像序列数据集上的有效性。

关键词: 图像序列, 谱流形学习, 混合多流形, 局部切空间, 层次流形