基于多示例学习的图像标注方法

Image Annotation Based on Multi-instance Learning Method

随着数码产品，移动智能设备以及存储设备的普及，大数据时代已经来临，如何对海量数据进行有效的组织、管理、存储成为科研以及商业领域急需解决的问题，在图像数据挖掘当中，图像标注分类是当前比较热门的方向。采用机器学习的方法来找到大规模数据当中的隐含规律，实现样本的视觉内容到概念的映射需要对视觉数据内容进行恰当的描述，如果我们使用整个的图像作为基本单元，存在的问题就是视觉数据往往具有歧义性，难以准确表述包含的语义，多示例学习方法应运而生。图像分类问题本身是一种多标签问题，传统方法将其转化为一系列的单标签问题解决，忽略了标签之间的相关性，我们将标签相关性引入到模型构建当中，实验取得良好效果。

With the popularity of digital products, mobile smart devices and storage devices big data era has arrived. It has became an urgent problem needed to solve both in research and commercial areas that how to organize manage and storage massive data which induced the research work of image data mining in which image annotation classification is a popular direction currently. In the work of using machine learning method to find hidden law in massive data we need to achieve the mapping from visual da-ta to concepts properly. If we use the whole image as basic unit there will be the problem of ambiguity and difficult to express the contained semantic labels correctly. In this kind of situation multi-instance leaning methods emerged. Image annotation is in-deed a multi-label problem and in traditionally it has been divided into several single label problem to deal with. In this kind of methods correlation between labels has been ignored, so we introduce label correlation into construction of model and it has achieved better results.