以內容為基礎的建築物影像檢索

Abstract

本論文目的在使用影像區域特徵來建立一個建築物影像檢索系統。此檢索系統分成資料庫與查詢兩個部份，資料庫部份按照處理順序又可分為三個步驟，第一步驟使用可抗視角變化的Maximally Stable Extremal Region做特徵區域擷取；第二步驟使用旋轉不變的phased-based Zernike Moment做特徵區域描述；第三步驟使用kd-tree建立特徵向量的索引。建立資料庫時，使用同一棟建築物相鄰的影像特徵互相比對，去除不穩定出現的特徵區域，並使用Density-Based Spatial Clustering of Applications with Noise分群法，以減少資料庫中存在的儲存重覆特徵問題。查詢部分採用kd-tree找最近點與鄰近點的便利性，以直觀的投票機制找出資料庫中與查詢影像最相似的建築物。

The goal of this thesis research is to construct a building image indexing and retrieval system. This system consists of two parts: the database organization (indexing) and the query part (retrieval). The database part is further composed of three modules. In the first module, view-invariant feature detection, Maximally Stable Extremal Region (MSER), is used to extract the regions of interest. In the second module, the phased-based Zernike Moment is used to describe the regions. In the third module, a kd-tree structure is used to establish the index of Zernike Moment feature vectors. When constructing the database, in order to eliminate the unstable regions, a trick of comparison of the features extracted from the neighboring views of the same building is used. To reduce the problem of redundancy, the clustering algorithm, Density-Based Spatial Clustering of Applications with Noise (DBSCAN), is used. In the query part, the kd-tree provides a convenient way to find the nearest neighbor. And then an intuitive voting mechanism is used to find the building from the database which is most similar to the query image.