基于相似度的两视角多示例图像分类方法研究

Research on Two-View Multi-Instance Image Classification Based on Similarity

在实际中,某些数据中包含许多特权信息,可用于训练分类器,从而提高分类性能。例如,在图像分类中,标签用于描述图像,这些标签可视为特权信息,特权信息与图像互补,可以用于学习以此提高图像分类性能。多示例学习和两视角学习的特性适用于带有特权信息的图像分类,因此提出了一种基于相似度的两视角多示例方法用于带有特权信息的图像分类。所提方法将一张图像视为一个示例,若干张图像的集合视为包,将特权信息视为示例。为解决实际中示例的标签是未知的问题,因而引入相似度模型。所提方法首先将图像和特权信息划分为两个不同的视角,然后使用聚类算法构造包,最后训练支持向量机分类器。在四个数据集上的实验结果表明,所提方法与其他相类似模型相比精确率更高,并比较了两种包的聚类算法,分析了各参数敏感度。

In practice, some data contains a lot of privileged information, which can be used to train the classifier to improve classification performance. For example, in image classification, labels are used to describe images. These labels can be regarded as privileged information. The privileged information is complementary to the image and can be used for learning to improve the performance of image classification. The characteristics of multi-instance learning and two-view learning are suitable for image classification with privileged information. Therefore, a two-view multi-instance method based on similarity is proposed for image classification with privileged information. The proposed method considers one image as an instance, a collection of several images as a package, and privileged information as an instance. In order to solve the problem that the labels in the instances are unknown in practice, a similarity model is introduced. The proposed method first divides the image and privilege information into two different perspectives, then uses a clustering algorithm to construct the package, and finally trains a support vector machine classifier. The experimental results on four data sets show that the proposed method is more accurate than other similar models, and the two packet clustering algorithms are compared, and the sensitivity of each parameter is analyzed.