一种新的基于多示例学习的场景分类方法

多示例学习是近年来才出现的一种新的学习框架,并以其对多义性对象的出色表示能力而被成功地运用在图像分类任务中。首先提出了一种新的图像多示例包生成方式,采用特征的概率分布表示图像,并对图像像素分布提取一个高斯混合模型,将每个高斯分布作为一个示例,生成图像的多示例包。然后,在对图像进行分类时,采用了信息瓶颈聚类把多示例包转化成单示例,从而将传统的单示例分类器用在该问题上。为了提高分类器的泛化能力,对多个分类器进行了集成。选取了5类自然场景图像进行试验,结果显示所提出的方法平均性能优于当前常用的一些多示例学习算法。

A new Multi-instance learning method for scene classification

Multi-Instance learning (MIL) is a learning framework proposed recently, and has been successfully used in scene classification. First,A new image Multi Instance (MI) bag generating method was proposed, which models an image with a Gaussian Mixed Model (GMM). The generated GMM was treated as an MI bag and the components of the GMM were the instances of the corresponding bag. Then, the information bottleneck clustering was employed to transform the MIL problem into single-instance learning problem so that single-instance classifiers could be used for classification. Ensemble learning was involved to further enhance classifiers’ generalization ability. Experimental results showed that the proposed method was superior to some common MI algorithms on average in a 5-class scene classification task.