图像分类中的概率乘积核函数

Probability product kernels in image classification

Bag-of-Features(BOF)模型使用编码矢量的某一特定统计值表征图像,与基于传统核函数的支持向量机相配合完成对图像的分类,所带来的问题是会丢弃大量判别信息以及最优核函数的选择。因此,本文将硬分配编码矢量服从的多项分布、软分配编码矢量服从Dirichlet分布,并以此作为图像的内容描述,利用最大似然算法估计其中参数,然后使用概率乘积核函数计算图像两两之间的核函数,最后使用支持向量机对图像进行分类。公开图像数据集上的实验结果表明,本文算法取得了更优的分类性能。

Images are characterized by some statistics of coded vectors in the Bag-of-Features （BOF）model, and then classified by support vector machine （SVM）based on traditional kernel, the existing problems are the loss of discriminant information and choosing of optimal kernel. To solve these problems, we use the multinomial distribution of hard coded vectors or Dirichlet distribution of soft coded vectors as the description of images, and then use maximum likelihood algorithm to estimate the density parameters. Next, the kernel functions between any two images are calculated using a probability product kernel function. Finally, the images are classified by a support vector machine. The experimental results in public image datasets show the proposed algorithm in this paper has achieved better classification performances.