为了有效地解决多示例图像自动分类问题,提出一种将多示例图像转化为包空间的单示例描述方法。该方法将图像视为包,图像中的区域视为包中的示例,根据具有相同视觉区域的样本都会聚集成一簇,用聚类算法为每类图像确定其特有的"视觉词汇",并利用负包示例标注确定的这一信息指导典型"视觉词汇"的选择;然后根据得到的"视觉词汇"构造一个新的空间——包空间,利用基于视觉词汇定义的非线性函数将多个示例描述的图像映射到包空间的一个点,变为单示例描述;最后利用标准的支持向量机进行监督学习,实现图像自动分类。在Corel图像库的图像数据集上进行对比实验,实验结果表明该算法具有良好的图像分类性能。

In order to effectively solve the multiple-instance image classification problem, we put forward a new classification method, which transforms the multiple-instance image into a single instance image in the new space-bag space. First, the whole image is regarded as a bag and each region as an instance of that bag. According to the same visual regions of image samples are put into one cluster and k-means clustering algorithm is used to determine the visual words for each class of images. At this step, we use the information that labels of negative samples are all known has been used to select the typical visual words. Then, we construct a new bag space with these visual words and use a nonlinear function based on these visual words to transform each multiple-instance image into a point in the bag space. Finally, standard SVMs are trained in the bag feature space to classify the images. Experimental results and comparisons on the Corel image set are given to illustrate the performance of the new method.