整合多搜尋方法之影像資料庫檢索系統

An Image Retrieval System Obtained by Integrating Multiple Searching Approaches

曾逸鴻(Yi-Hong Tseng) ； 黃吉緯(Chi-Wei Huang)

《科學與工程技術期刊》 5卷3期 (2009/09) Pp. 1-12

摘要

現存的影像檢索系統，大多是針對特定影像資料庫，抽取適當特徵值用以尋找相近的影像；然而，同樣的搜尋方法不一定適用於其他類型資料庫。以顏色特徵值爲基礎的搜尋方法，分別應用在黑白、灰階及彩色影像資料庫中，檢索效果有明顯的不同；而以形狀特徵值爲基礎者，則適用於前景物體存在且和背景環境差異分明的影像資料庫。有時，使用者又希望直接以影像內容的主觀描述（如影像分類、標的類型、拍攝環境等）來檢索相關影像。因此，本研究分別以影像的顏色特徵、形狀特徵與內容特性，來建構不同的影像搜尋方法。在影像資料庫建構過程，所有影像利用改良式K-means演算法進行分群，並於分析各群集中影像特徵之變異程度，於檢索時給予各個群集不同權重，開發出一套整合多搜尋方法並適用不同影像資料庫之檢索系統。經過完整實驗步驟之測試，可顯示整合三種搜尋方法，所得之影像檢索效果最佳。

關鍵字

多搜尋方法 ； 影像檢索系統 ； 特徵分群

並列摘要

Extant image retrieval systems are mainly aimed at specific databases which extract and match appropriate features to retrieve similar images. However, the same searching approach cannot be successfully applied in other kinds of databases. To search for images from a binary, gray or color-image database by matching color features results in a discrepant retrieval. Several shape-based searching methods are usually used to retrieve images in which foreground objects are distinct from the background environment. Sometimes users want to search for images by inputting subjective descriptions such as the type of image, object or pictorial environment. Hence, this research implements different searching modules by considering the color-, shape-, and content-based characteristics of various images. During the construction of an image database, all items are aggregated into feature-clusters by a modified K-means algorithm. Different weights are then assigned to each cluster by analyzing the characteristic variations in all images in a particular group. Thus, in this study, an image retrieval system obtained by integrating multiple searching approaches is ultimately achieved. After the experimental steps were completed, the results showed that the best retrieval effect is obtained by integrating the three aforementioned searching approaches.

並列關鍵字

multiple searching approaches ； image retrieval system ； feature-clustering