累積由過往之查詢期間(Query Session)所學習到的知識用以增進以內容為基礎的影像資料庫檢索 (Content-Based Image Retrieval)之檢索效果是目前影像檢索研究的主要課題之一。經由查詢期間歷史(Query Session History)所得到的資訊可以有效的克服語意隔閡 (Semantic Gap)並顯著的增進檢索效果與效率。本篇論文提出一個能以隱含語意空間(Hidden Semantic Space)累計過去查詢經驗的影像檢索系統。首先,在每次的查詢期間根據使用者提供之相關回饋(Relevance Feedback)學習出一個支援向量分類器,並以該分類器估算影像含有該查詢期間所代表之隱含語意的機率。以多次查詢期間所習得之隱含語意機率建構出隱含語意空間,在往後的查詢期間便可用此隱含語意空間所提供之隱含語意特徵進行檢索。為了避免隱含語意空間因持續累積查詢期間的資訊而造成空間維度過大,我們提出以圖像為基礎之群聚整合學習的方法根據現有的隱含語意特徵對影像進行分群以縮減隱含語意空間的大小。分群後之每個群聚即為一個整合過之隱含語意概念。經由設計的實驗,我們可證實我們所提出的以圖像為基礎之群聚整合學習方法能有效率的降低隱含語意之維度,並保持隱含語意空間的有效性及可靠性且提升檢索系統的效能表現。
Inter-session learning in content-base image retrieval (CBIR) makes user take advantage of the information learned from previous query session. Many works have been proposed for the inter-session learning. In this thesis, the basis is a framework using a hidden semantic space to accumulate the inter-session information. At first, we use the SVM classifiers trained in short-term learning to initialize the hidden semantic space with a probabilistic model. To maintain the hidden semantic space in a proper size, we propose a novel framework based on graph-based cluster ensemble. Each time the hidden semantic space is over-expanding, we use our proposed dimension reducing method to construct a compact, effective, meaningful, and lower dimensional hidden semantic space. With the hidden semantic space, a long-term learning scheme is performed. Our experimental results demonstrate that the graph-based cluster ensemble scheme works well and efficient in our long-term learning CBIR system. The scheme takes short time but provides stable and reliable results.