基于概率密度分布一致约束的最小最大概率机图像分类算法

Image classification algorithm based on minimax probability machine with regularized probability density concensus

为了解决含有大量未标记样本的图像分类问题,提出了基于概率密度分布一致约束的最小最大概率机图像分类算法(image classification algorithm based on minimax probability machine regularized by probability density concensus,PDM PM)。用概率密度估计函数对标记图像样本和未标记图像样本在超平面所在空间的分布进行估计,最小化标记图像样本和未标记图像样本在超平面所在空间的分布差异,得到概率密度估计约束项。把概率密度估计约束项融入到非线性最小最大概率机并用于图像分类。在耶鲁大学人脸数据库、加利福尼亚理工学院101类图像数据库的5类和15场景数据库中的10类分类准确率的试验中,该算法较非线性最小最大概率机大约平均提高了3.99%,从而表明该方法充分利用了未标记图像样本包含的分布信息来约束标记图像样本的分布,使得图像分类超平面更加准确。

In order to solve image classification problem of which the images contained labeled and unlabeled samples,this research proposed an image classification algorithm based on minimax probability machine regularized by probability density concensus（ called PDM PM）. The distribution of the image samples in the hyperplane was estimated by using the probability density estimation function and probability density estimation constrained item was got by minimizing the distribution of the labeled and unlabeled samples. The probability density estimation constraint item was integrated into the nonlinear minimax maximum probability machine and used for image classification. The accuracy of the proposed algorithm was increased by 3. 99% compared with Gaussian kernel minimax probability machine in the test of Yale face database,five of the Caltech 101 database and ten of Fifteen Scene Categories Dataset. Experimental results indicated that the method made full use of the distribution information of unlabeled image samples and made the image classification hyperplane more accurate.