摘要： 提出一种基于类别相关码本的方法，该方法为每一个类别产生一个码本，在训练和测试任意2个类别之间的分类器时，仅考虑与这2类相关码本上形成的图像向量。在保留相关类别码本多样性的同时，能降低输入分类器的图像向量的维数，有效避免维数灾难。实验结果表明，与传统的基于单个全局码本的方法相比，该方法具有更好的分类性能。

关键词: 类别相关, 码本生成, 图像分类, 图像向量化

Abstract: This paper proposes a strategy which generates a codebook for each class. When the classifier between any two classes is trained, only the image vectors generated from the codebooks related to these two classes are considered. Compared with the traditional method using one universal codebook, the dimension of image vectors is greatly reduced while the diversity of them is preserved in the approach, which avoids the occurrence of the curse of dimensionality. Experimental results show that the approach can gain better results than traditional methods which use only one codebook.

Key words: class-specific, codebook generation, image classification, image vectorization

中图分类号: 

• TP391.41