针对目前词袋(BoF)特征压缩算法忽略编码矢量之间空间关系的问题,本文给出了压缩算法与金字塔模型相配合的图像分类步骤。同时以多个公开图像数据集为实验对象,对典型词袋特征压缩算法的性能进行比较性研究报道。实验结果表明,压缩算法对于视觉单词数目以及编码方法具有良好的鲁棒性;其中基于子空间方法的压缩算法在高层图像特征空间中的分类性能最优,在多个图像数据集上的分类性能最优,时间开销最小。

Against the problem that compression algorithms for bag-of-features(BoF) ignore the spatial relationships of coded vectors, we propose a fusing algorithm of compression algorithms and spatial pyramid model in this paper. Meanwhile, we carried out a set of comparative experiments on several public image datasets. The experimental results show that compression algorithms are robust to visual word numbers and pooling methods of coded vectors. Otherwise, compression algorithms based on subspace methods have achieved best classification performances in the high-level feature space, and best accuracies and smallest time cost in multiple image datasets.