摘要： 跨数据集图像分类是在图像分类应用中经常面临的问题。由于训练集数据与实际待分类(或测试)数据既有内在联系,又具有较大差异,导致常见分类器在跨数据集分类中的性能明显下降。为此,根据数据信息提出一种新的跨数据集图像分类方法。将主成分分析中特征信息保留的思想引入到基于L1特征选取的Logistic回归中,从而在选取图像特征时有效保持数据集中的高信息含量特征。实验结果表明,在多个常用跨数据集图像分类中,该方法能获得较好的图像分类效果。

关键词: 图像分类, 跨数据集, 特征选择, Logistic回归, 稀疏主成分分析, 转换学习

Abstract: Cross-dataset image classification is a common problem in the real applications of image classification.Even though training data and testing data are related in the cross-domain classification,there are some differences between them.And this leads the performance of traditional classifier in cross-dataset classification dramatically reduced.In order to solve this problem,this paper proposes a novel cross-dataset image classification method.The new method introduces the idea of feature information reservation of Principal Component Analysis(PCA) into Logistic return based on L1 logistic regression,so that it can keep high information features in dataset when selecting image features.Experimental results show that in commonly used cross-dataset image classification,the method can obtain good image classification effect.

Key words: image classification, cross-dataset, feature selection, Logistic regression, sparse Principal Component Analysis(PCA), transformative learning

中图分类号: 

• TP391