一种基于图像集的复杂场景分类方法

Image Set Based Scene Image Classification

摘要:

保持局部图嵌入的流形鉴别分析方法将图像集所属子空间看作流形上的点，并使流形变换前后局部结构关系不变。然而在构造局部区域相似图矩阵时，用于描述节点局部区域范围的近邻节点个数会极大的影响算法的准确率，并会出现变换后流形的可分辨性相比变换前提升很小甚至更低的情况。针对该问题，提出了一种低秩描述下的Grassmannian流形鉴别分析方法。通过对图像集的低秩描述，流形变换中局部嵌入时仅保持同类别节点的最近邻局部结构以及所有节点间的相异类别信息，从而避免了对近邻节点个数的选择，并增强了变换后流形的可分辨性。在15类复杂自然场景和Caltech101图像数据集的实验结果表明该方法是可行的并且极大的提高了图像集分类的准确率。

Abstract:

Graph embedding discriminant analysis on manifold approach represents each image set as a subspace on manifold. It maps the manifold to a more discriminative one with geometrical structure and local information preserved. However, its accuracy critically depends on the number of local neighbours when constructing similarity graph. In this paper, we present a novel approach with fixed neighbour numbers to implement graph embedding Grassmannian discriminant analysis based on low-rank representation (LRR) for each image set. After recovering the low-rank components of each set, we find that preserving the nearest neighbour structure of nodes with same label and all the different label information during the manifold mapping can always achieve the best performance. Experiments on two image datasets (15-scenes categories and Caltech101) show that the proposed method greatly improves the classification accuracy of image sets.