內容式影像檢索是一種萃取影像低階內容，例如顏色、花紋、形狀等進行影像資料庫索引與查詢。然而，內容式影像檢索的主要問題在於萃取的影像低階特徵並不能完整的代表使用者的高階語意查詢。影像註解是目前解決此語意鴻溝問題的熱門方法，它主要是使用機器學習技術來自動學習影像低階特徵進而辨識影像內容而得以下註解。目前的文獻大多是致力於提出新的學習技術並提高影像註解的正確率進而提高後續的查詢效能。然而，除了使用低階特徵外，目前鮮少有研究著重在萃取更好與更有解釋能力的特徵進行影像註解。本研究提出一個新的後設特徵向量空間模型基於萃取影像資料與其分群中心及最近鄰居的一些相信距離資訊。總共有九種後設特徵的值。而我們認為這些資訊(即後設特徵)能提供更有區別各種不同之影像高階內容的能力。本研究主要有兩個研究目的。第一個目的是建構此後設特徵向量空間模型以 Corel 資料集為例(此為影像檢索領域最常使用的資料集)，而且總體與局部的影像內容皆會用來萃取後設特徵進行比較。此目的是用以了解後設特徵是否能針對總體與局部的影像註解都能提高正確率。第二個目的是進一步的檢視後設特徵的可應用性，也就是使用ImageCLEF 與TRECVID 資料集進行實驗，分別為醫療影像與視訊檢索的問題領域。在分類器製作方面，k 最接近鄰居、支援向量機與貝氏分類器將會用來比較後設特徵於影像註解的正確率。 level features), such as colour, texture, and shape to search images. However, the major limitation of CBIR is that the low-level features of images for indexing and retrieval cannot directly correspond to the high-level concepts of users’ need and requirements. Image annotation is one important solution to bridge the semantic gap problem of CBIR. It focuses on using some machine learning techniques to automatically learn the visual contents to recognize (such as classify and cluster) the semantic content of images. In literature, many novel machine learning techniques have been proposed to solve the semantic gap problem, whose aim is to improve the image annotation performance for better retrieval effectiveness. However, related studies rarely focus on extracting better features in addition to using low-level image features directly for image annotation. This project proposes a novel meta-feature vector space model by extracting some distance information between the image data (represented by low-level feature vectors) and their cluster centers and nearest neighbors. In total, there are nine meta-features, which can be extracted. This kind of information is supposed to provide more discrimination powers for effective image annotation. There are two research objectives of this research project. The first objective is to develop the meta-feature vector space model for image annotation over the Corel dataset (the most popular one in image retrieval). Particularly, the meta-feature extracted from global and local image contents are assessed in terms of image annotation accuracy. The second objective is to further examine the applicability of the meta-feature over ImageCLEF and TRECVID datasets, which belong to the medical images and video retrieval problems. On the other hand, k-nearest neighbor, support vector machines, and naïve Bayes are used as the classifiers for image annotation. 研究期間：10008 ~ 10107
