9D-SPA+: 影像資料庫中空間關係之知識表示法

Abstract

在本篇論文中，我們提出一個新的空間關係之知識表示法「9D-SPA+」來表示一張符號圖像。植基於9D-SPA+，我們可以擷取物件間完整且正確之空間關係，同時也保留了物件的長寬與物件間距離的資訊。此外，我們針對9D-SPA+提出一個影像重構的演算法，藉此將9D-SPA+表示法轉換成符號圖像以便瀏覽。為了獲得圖像中物件間的順序與距離，我們首先將9D-SPA+表示法進行解碼，並建構一個空間關係圖形。藉由此圖形，我們可以輕易地重建出原始的符號圖像。最後，我們提出一個彈性的相似度比對演算法用以搜尋圖像資料庫，其利用物件間的空間關係和物件的長寬資訊，來評估查詢圖像與資料庫圖像的差異。藉由調整不同比對方法的權重，我們提出的方法可以滿足不同使用者的需求。最後的實驗結果證明，我們所提出的方法具有效率及擴充性，且我們所提出的多種相似度比對方法，對於不同種類的圖像皆具有高度的辨識能力。

In this thesis, we have proposed a new spatial knowledge structure, called 9D-SPA+, to represent symbolic images. Based on the 9D-SPA+ knowledge structure, we can capture the precise and compact spatial relations between objects and preserve the metric information of objects without ambiguity. Moreover, we propose an image reconstruction algorithm for the 9D-SPA+, which converts a 9D-SPA+ representation into a symbolic image for visualization and browsing. In order to obtain the orders and distances among objects in an image, we decode the 9D-SPA+ representation first and construct spatial relation graphs for reconstruction. Based on the graphs, we can easily reconstruct a symbolic image for a 9D-SPA+ representation. Finally, we present a flexible similarity retrieval algorithm for retrieving the similar images from the image database, which infers spatial relations among objects and metric information of objects to assess the difference between the query and database images. By adjusting the weights of different measures, our proposed method can meet users’ requirements. The experiment results show our proposed approach is scalable and efficient. Furthermore, by providing various measures for similarity retrieval, the experiments demonstrate our proposed similarity retrieval algorithm has discrimination power about different criteria.