Region-Based Image Retrieval by Use of Relevance Feedback

由於影像的資料庫越來越大，有效的處理大量的資料庫成為一項重要的課題。本篇論文使用了以區域為基礎的影像檢索技術，建立一個影像檢索系統。與一般影像檢索不同的地方在於使用者可以自由的選擇影像中特定的區域，再從資料庫中找到與這個特定區域相似的圖片。由於每張影像還要切割成一些小區域，因此整個資料庫將會建立很多區域的資訊，在此我們提出了一個改良式區域分類技術來幫區域分類，並且利用它把資料庫內相似度太低的影像過濾掉。而剩下那些被認為相似度較高的影像，我們再給他排名找出最相似的影像。另外本篇論文還加入了關聯回饋技術，透過關聯回饋，我們可以清楚地知道使用者所與查詢的概念，進而找出更符合使用者所想要的影像。With the exponential growth of multi-media data, finding images in a large database has become more difficult. Region-based image retrieval (RBIR) is used for solving this problem in this thesis. There are some differences between RBIR and traditional content-based image retrieval (CBIR) systems. CBIR is focused on the similarity of global images and RBIR is focused on the similarity of the local image regions. We apply the watershed segmentation to segment each image into some regions. To classify these regions, the fuzzy k-means clustering algorithm is time-wasting and uses too much space to store the information about the regions. We propose a modified fuzzy k-means clustering algorithm to classify regions efficiently. In order to accelerate our system, we propose a new method for filtering which can filter out many unsuitable images. The candidate images are ranked based on their similarity measure. After our system retrieves the images, the user is able to give feedback to the system. Based on user’s feedback information, our system will retrieve the images that are even closer to the user’s intent.Chapter 1 1 Introduction 1 1.1 Content-Based Image Retrieval 1 1.2 Region-Based Image Retrieval 4 1.3 Relevance Feedback 6 1.4 Thesis Organization 7 Chapter 2 8 Background 8 2.1 Watershed Segmentation 9 2.2 Feature Extraction 11 2.2.1 Color Feature 12 2.2.2 Texture Feature 16 2.3 Clustering 18 2.3.1 k-Means Clustering 19 2.3.2 Fuzzy k-Means Clustering 20 2.3.3 Adaptive Clustering 22 2.4 Earth Mover’s Distance 23 Chapter 3 29 The Proposed Method 29 3.1 Image Representation 29 3.1.1 Modified Fuzzy K-means Clustering 31 3.1.2 Indexing and Image Representation 32 3.2 Image Filtering 33 3.3 Image Ranking and Retrieval 34 3.4 Relevance Feedback 36 Chapter 4 39 Experiments 39 4.1 Image Database 39 4.2 Experimental Results 40 Chapter 5 46 Conclusion and Future Work 46 5.1 Conclusion 46 5.2 Future Work 4