基于黎曼字典和ScSPM模型的场景分类

Scene classification based on Riemannian dictionary and ScSPM model

场景图片的局部信息包含2类:像素统计分布信息和像素动态变化信息.经典场景分类模型Sc SPM仅考虑了前者,利用协方差矩阵度量局部像素动态信息,作为Sc SPM模型的互补分类信息,以提高场景分类准确率.为解决协方差矩阵特征的聚类问题,使用基于黎曼度量的顺序聚类算法,然后利用提出的局部软分配算法编码协方差矩阵.标准场景库的实验证明了所提思路能够显著提升Sc SPM模型的分类精度.

Scene images contain two kinds of information: statistical information and dynamic information of local pixels. The classical scene classification model ScSPM only focuses on the former's information. This paper presents an improved ScSPM model based on the incorporation of a Riemannian dictionary of covariance descriptors. In the proposed approach,Riemannian dictionary over the descriptors is built by the recursive mean algorithm,and the revised version of localized soft-assignment coding method is proposed to encode the descriptors. The experimental results show that the improved ScSPM outperforms state-of-the-art methods on the LS,and achieves the competitive performance.