一种通过去噪来提高图像标注性能的方法

提出一种通过去除噪声来提高分层高斯混合模型的图像标注性能的方法．在有监督多概念标注（supervised Multiclass Labeling，SML）中，它采用了分层高斯混合模型，有效地把大量的表示图片的高斯混合模型聚类成一个表示某概念类的高斯混合模型，从而很容易实现图像的分类和标注．通过引入去噪过程，减少了图片中与概念不相关的高斯分量，消除了噪声高斯分量对分类器的负面影响，从而提高了分类性能．基于TRECVID2005数据的图像标注实验证明，该方法可以取得比SML方法更好的标注性能．

Imroving concept modeling through denoising for image annotation

We present a clustering method of improving image annotation performance by denosing.Similar to the Superviese Multiclass Labeling(SML) method,it adopts Hierarchical Gaussian Mixture Model(HGMM) which is efficient in reducing a large mixture of Gaussians into a smaller number of mixture components.Compared with SML method,our method reduces the noisy(i.e.irrelevant) component of GMM so as to eliminate their negative effects for HGMM classifiers,thus improves the performance of concept modeling.Image annotat...