Abstract: An improved Bag Of Visual Words（BOVW） representation algorithm based on maximum frequent item-sets is proposed. Isolated points are ruled out and an efficient mining of maximum frequent item-sets based on annular region division is used to find visual words occurring frequently. The proposed algorithm highlights the differential features between different categories and spatial information is contained. In comparison, traditional BOVW could not fully express image common characteristics on one category. Experimental results on COREL and Caltech-256 database demonstrate the effectiveness and feasibility of proposed algorithm.

Key words: image categorization, visual words, maximum frequent sets

摘要： 针对传统视觉词袋（Bag Of Visual Words，BOVW）模型缺少空间信息，且不能充分表达图像所属类别共有特征的问题，提出一种基于最大频繁项集的视觉词袋表示方法。该方法在排除孤立特征点的基础上，引入环形区域划分的思想，嵌入更多的空间信息。通过对不同环的视觉单词进行频繁项挖掘得到新的视觉单词表示，能有效提高同类别图像视觉单词的相似程度，而使不同类别视觉单词的差异更为显著。通过在图像数据集COREL及Caltech-256上进行分类实验，验证了该方法的有效性和可行性。

关键词: 图像分类, 视觉单词, 最大频繁项集