基于成分金字塔匹配的对象分类方法

提出基于成分金字塔匹配(component pyramid matching, CPM)的图像表示方法,将图像块按照颜色进行分层,在每一层中通过优化的方式选取几种颜色的图像块作为当前层次图像的前景成分,其余颜色的图形块作为图像的背景成分。前景成分对应对象的某些区域,能够为图像表示提供弱语义信息。然后,利用相似的颜色选择方法,对每一层背景成分进行再次划分,将其分为下一层前景成分和背景成分两部分。最后将这些成分所表示的直方图连接起来作为图像表示用于分类。试验采用Soccer、Flower17和Flower102 3个图像集进行测评,试验结果表明提出的算法能够得到比较好的分类结果。

Object classification method based on component pyramid matching

The image representation method based on component pyramid matching(CPM)was proposed, which separated the patches into different levels based on colors. In each level, some colors were selected by the optimal color selection method, then the patches with these selected colors were considered as the foreground components, and the rest of the patches with other colors were considered as the background components. Usually, the foreground components corresponded to some parts of the objects, which could supply weak semantic information for the image representation. Then, the background components were split into the foreground and background components in the next level based on the similar color selection method. The final representation of an image was obtained by concatenating the component histograms in each level. Classification results were presented on Soccer, Flower17 and Flower102 datasets, and the experiments showed that CMP could obtain satisfactory results in these datasets.