- 學位論文
一個通用有效的兩階段區塊式影像檢索方法
A General and Effective Two-Stage Approach for Region-Based Image Retrieval
摘要
過去數十年間，隨著多媒體應用和數位典藏的快速發展，內涵式影像檢索(Content-Based Image Retrieval, CBIR) 已經吸引了許多關注且演變成為一個重要的研究領域。內涵式影像檢索多以影像的低階內容來建立索引和進行檢索，由於它在整個過程中牽涉各式各樣的議題，如色彩空間的選擇、特徵擷取、相似度測量、檢索策略、相關度回饋等，使得它成為一個既複雜又具挑戰性的問題。這些議題中，“語意差距” (Semantic Gap) 仍然是內涵式影像檢索中極富爭議與挑戰的問題，它反映了利用低階特徵來檢索與使用者所需的高階意涵兩者間所存在的不一致性。有一些學者嘗試利用區域特徵來縮小這個差距，也就是所謂的區塊式影像檢索(Region-Based Image Retrieval, RBIR)。 區塊式影像檢索主要是搜尋與檢索目標相近的區域，而不是整個影像；它提供了一個比較有意義的影像檢索。然而，影像分割演算法不但複雜、計算耗時且其分割結果往往不正確。為了解決這個問題，我們提出了一個兩階段的檢索策略來改進區塊式影像檢索的效能。在第一階段， 我們使用以門檻為基礎的刪除法(Threshold-Based Pruning, TBP)來當作一個過濾器，快速刪除那些整體特徵差異極大的影像；在第二階段，一個細部特徵比較法(Detailed Feature Comparison, DFC)則被用來檢驗剩餘的候選影像與感興趣區域(Region of Interest, ROI)之間的差異。在我們的實驗系統中，使用者可以在查詢影像中選擇他們的感興趣區域，並在檢索過程中與系統進行互動，如選擇不同的檢索策略、設定參數值、調整特徵的權重等。實驗結果顯示我們所提出的兩階段檢索方法分別改進了影像檢索的效率和正確性達10.7%與7.1%。
並列摘要
With the rapid growth of multimedia applications and digital archives, content-based image retrieval (CBIR) has received lots of attentions and emerged as an important research area for the past decades. CBIR tends to automatically index and retrieve images based on their low-level contents, which is a complex and challenging problem spanning diverse algorithms all over the retrieval processes including color space selection, feature extraction, similarity measurement, retrieval strategies, relevance feedback, etc. In these issues, “semantic gap” is still an open challenging problem in CBIR. It reflects the discrepancy between low-level features developed by the retrieval algorithm and high-level concepts required by users. Some research works attempt to narrow this gap by utilizing regional features, which are known as region-based image retrieval (RBIR). RBIR tends to search the interesting regions that are closed to the query target, instead of the whole images. It contributes to more meaningful image retrieval; however, the image segmentation algorithms are complex and computation intensive and the segmentation results are often not correct. To tackle this problem, we propose a two-stage retrieval strategy to improve the performance of RBIR. At the first stage of retrieval, the threshold-based pruning (TBP) serves as a filter to quickly remove those candidates with widely distinct global features. At the second stage, a more detailed feature comparison (DFC) method is conducted over the remaining candidates, focusing on the region of interest (ROI). In the experimental system, users can choose their ROI in the query image and interact with the system by selecting different strategies, setting parameter values, and adjusting the weights of features as the search progresses. The experimental results show that both efficiency and accuracy can be respectively improved by 10.7% and 7.1% using the proposed two-stage approach.