Adaptive Content-based Image Retrieval using Semi-Supervised Learning Method

图像自动标注是图像数据库检索的一个很有希望的子技术。我们一直在构建一个基于半监督学习方法的图像自动标注产生式模型系统。由于它对于高维很容易不稳定，所以我们必须预先应用降维方法。通常，传统的监督降维方法(使用标记样本)在小样本情况下存在退化协方差矩阵问题。另一方面，无监督降维方法(使用未标记样本)不能正确识别类别之间的差异。在这项研究中，我们提出了一种新的半监督降维方法，该方法使用少量的标记样本和大量的非标记样本。实验结果表明，该方法的分类正确率比无监督方法提高了5.1个百分点。

Automatic image annotation is a hopeful sub-technique for image database retrieval. We have been constructing a generative model system for automatic image annotation using semi-supervised learning method. As it can be easily unstable for the higher dimensions, we must apply a dimensionality reduction method in advance. Generally, conventional supervised dimensionality reduction method (using labeled samples) suffers from the degenerate covariance matrix problem in the case of a small number of samples. On the other hand, unsupervised dimensionality reduction method (using unlabeled samples) can't recognize the differences among the categories properly. In this study, we propose a novel semi-supervised dimensionality reduction method using a small number of labeled samples and a large number of unlabeled samples. By the result of experiments, the classification rate of the proposed method was 5.1 points better than that of the unsupervised method.