摘要： 传统词包(BOW)模型中的视觉单词是通过无监督聚类图像块的特征向量得到的，没有考虑视觉单词的语义信息和语义性质。为解决该问题，提出一种基于文本分类的视觉单词歧义性分析方法。利用传统BOW模型生成初始视觉单词词汇表，使用文档频率、χ2分布和信息增益这3种文本分类方法分析单词语义性质，剔除具有低类别信息的歧义性单词，并采用支持向量机分类器实现图像分类。实验结果表明，该方法具有较高的分类精度。

关键词: 图像分类, 视觉单词, 文本分类, 支持向量机, 词包模型

Abstract: Visual words in the traditional Bag of Word(BOW) model can be gotten by an unsupervised method of clustering the visual features. But one critical limitation of existing BOW is not concerned with the semantic natures of visual words. This paper proposes a visual words ambiguity analysis method based on text categorization. The codebook is generated by the BOW model. There are three ways of analysis——document frequency, χ2 distribution and information gains, and then they reduce the low information visual words after analyzing. It gets optimized visual words, the histogram formed by the frequency of visual words is used in image categorization task by the Support Vector Machine(SVM) classifier. Experimental results show that this method has higher classification accuracy.

Key words: image classification, visual words, text classification, Support Vector Machine(SVM), Bag of Word(BOW) model

中图分类号: 

• TP306