利用频繁模式挖掘进行图像标注

Using Frequent Pattern Mining for Image Labeling

在基于内容的图像检索与计算机视觉研究领域中,如何将底层的视觉特征与高层的语义信息相联系,即如何有效地根据图像的底层特征提取其表达的语义概念是备受关注的难题之一。特别是当图像包含了多个语义概念时,问题就变得更为棘手了。本文中,我们提出一种基于图像底层特征值频繁模式的语义概念标注方法,针对图像分块的特点实现了一组有效的模式挖掘算法,并设计了标注规则的生成算法。权威的真实数据集上的实验表明我们的方法在对含有多个语义概念的图像进行概念标注时要比之前的一些算法效果更好。

One major challenge in the content-based image retrieval and computer vision research is to relate low-level visual features with semantic concepts, that is, to extract semantic concepts from an image effectively according to its low-level visual features. Especially when images contain more than one concept, the problem will be even more intractable. In this paper, we provide a method to label an image based on the frequent patterns of its low-level visual features. According to the specialty of image segmentation, effectivealgorithms are implemented to mine such patterns and to generate labeling rules. It is shown in the experiments on authoritative and real datasets that our method is more effective than some previously proposed methods for labeling images containing multiple concepts.