Abstract: Similarity measurement of image classification leads to outliers caused by vague boundary of feature points. To address this problem, an image classification method based on feature point class separability criterion is proposed. This paper combines information entropy theory used to extract the separability characteristics of image features. The separability distance values between feature vectors are calculated, and the nearest neighbor feature vector set is obtained according to the different nature of marking decision attribute by feature vectors. A new image type criterion is defined from two ways, the one is average distance between every feature vector of test image and the marking image class by nearest neighbor feature vector set, the other is average separability measurement value. The theoretical analysis and the Caltech256 simulation experiment show that compared with other methods, the feature point class separability criterion improves the image classification accuracy effectively.

Key words: image classification, separability, distance measurement, decision attribute, nearest neighbor classification

摘要： 针对图像分类特征点特性界定模糊，导致相似性度量误差较大的问题，提出采用特征点类别可分性判断准则的图像分类方法。结合信息熵理论提取图像特征点的可分性特性，根据图像特征向量标识决策属性的不同性质，计算特征向量间的可分性距离值，得到最近邻特征向量集，从待分图像各特征向量与最近邻特征向量集标识类别的平均距离，及平均可分性度量值两方面定义新的图像类别判断准则。理论分析与Caltech256图像库仿真实验表明，基于特征点类别可分性判断准则有效地提高了图像的分类准确率。

关键词: 图像分类, 可分性, 距离度量, 决策属性, 最近邻分类