中文摘要：

在高斯图特征提取过程中，通用背景模型（Universal background model，UBM）方法常用于根据总体分布估计每一幅图像中特征点分布的高斯混合模型（Gaussian mixturemodel，GMM）参数．然而UBM估计的GMM权重参数中有很多接近零的数值，它们所对应的高斯分量对分布估计贡献小却又都参与了计算，因此UBM的时间复杂度较高．为解决这个问题，本文提出BayesUBM方法．通过引入受限的对称Dirichlet分布来描述GMM权重参数的先验分布，利用Bayes最大后验概率对GMM参数集进行估计．实验表明BayesUBM方法不仅有效地降低了时间复杂度，而且提高了Corel数据集上的图像标注精度．

英文摘要：

The universal background model （UBM） is commonly used for Gaussian map feature extraction. The UBM estimates the parameters in the Gaussian mixture model （GMM）. However, the weight coefficients of GMM estimated by UBM have many near-zero values, whose corresponding Gaussian components have little contribution to the estimated result but need to be calculated in model estimation, therefore, UBM has a high time complexity. To solve this problem, we propose a method called Bayes UBM. In this method, the symmetric Dirichlet distribution is introduced to describe the prior distribution of GMM weight coefficients. The posterior distribution of the GMM weight coefficients is computed using Bayes method to estimate the GMM parameters. Experiments show that the proposed Bayes UBM method can efficiently reduce the time complexity, and improve the image annotation precision on Corel dataset.