- 學位論文
以影像特徵為基之圖片知識搜尋技術
A Picture Knowledge Extraction Methodology Based on Image Features
摘要
在日常生活中，人們常可能對陌生圖片產生興趣，並欲進一步瞭解此些圖片內容之相關資訊（即圖片知識）。而當人們欲搜尋其感興趣圖片（即目標圖片）之相關知識時，往往需先釐清與該圖片相關之關鍵字，再以此些關鍵字為基礎，進一步藉由文字型搜尋引擎搜尋此些圖片之相關知識。但人們卻往往難以由陌生之圖片釐清相關之關鍵字，造成其不易準確地搜尋此些圖片之相關知識；此外，既有搜尋引擎之搜尋結果大多以雜亂之條列方式呈現，使得目標圖片相關之重點知識常零碎地分佈於各筆搜尋結果中，造成搜尋者無法有系統地透過既有搜尋引擎之搜尋結果取得其期望之完整圖片知識。故本研究乃提出一套「圖片知識搜尋」方法論，以讓知識吸收者可直接藉由其感興趣之圖片搜尋與圖片內容相關且完整之知識，即使得知識吸收者更加有效率地取得其感興趣之圖片知識，並可獲取更完整、具系統性之目標圖片相關知識。 發展此「圖片知識搜尋」方法論前，本研究乃先歸納各類型圖片之圖片知識表達內容，以將圖片知識內容分類並結構化，並根據此結構化結果整理樣本圖片之相關知識；其次，再評估圖片中對各類圖片知識較具影響性之影像特徵，進而釐清各影像特徵對各類圖片知識之相關程度，作為「圖片知識搜尋」方法論進行圖片知識擷取之基礎。之後，本研究即發展一套可協助知識吸收者以感興趣目標圖片搜尋相關圖片知識之「圖片知識搜尋」方法論，而本方法論之詳細步驟乃先取得知識吸收者感興趣目標圖片之影像特徵，並根據此些影像特徵將目標圖片與基準圖形知識庫之樣本圖片進行比對，以取得基準圖形知識庫中與目標圖片具高度關聯性之樣本圖片群，並將此些樣本圖片之結構化圖片知識加以篩選，而形成目標圖片之各類相關知識，再根據先前釐清之影像特徵對各類圖片知識相關程度調整各類知識內容之排序，進而提供一份較完整且具有系統之目標圖片相關知識予知識吸收者參考。最後，本研究亦根據此方法論建構一套「圖片知識搜尋」系統，並以「博物館館藏品圖片」作為應用案例，以評估本系統之圖片知識推論績效。而由驗證結果分析可知，本研究所開發之「圖片知識搜尋系統」可有效地應用於實際之圖片知識搜尋情境，且其效能頗佳。整體而言，本研究所提出之圖片知識搜尋技術可促使圖片蘊含知識更能被知識吸收者準確查詢與應用，進而提升圖片之再利用價值。
並列摘要
In our daily life, we might be interested in the implicit knowledge (i.e., picture knowledge) behind some unknown pictures (namely the target pictures). Traditionally, in order to search the picture knowledge via a search engine, users have to identify some keywords related to the target picture. It is not always easy for users to identify keywords to accurately represent the target picture and as a result the users cannot accurately acquire the implicit knowledge regarding the target picture. Furthermore, the search results via a search engine are not well-organized and the picture knowledge might be fragmented in distinct search results. This research aims at developing a feature-based picture knowledge extraction methodology to assist the knowledge acquirers to extract the picture knowledge via the target picture they are interested in and acquire a more completed and systematic search result. At first, this research analyzes the picture knowledge in different application domains, and identifies the components of picture knowledge. Based on the knowledge components, this research establishes a sample picture database and reorganizes the knowledge of these sample pictures. After that, the key image features for picture knowledge extraction are identified and a feature-based picture knowledge extraction methodology is developed. In the proposed methodology, after acquiring the target picture from the knowledge acquirer, a set of sample pictures highly correlated with the target picture can be determined based on the image features for the target picture and sample pictures. After that, the knowledge regarding the extracted sample pictures are integrated to represent the knowledge of the target picture. Based on the proposed methodology, this research also develops a picture knowledge extraction system. Furthermore, the cultural relic pictures are employed as examples to evaluate the proposed methodology. The verification results show that the developed system can be applied to real cases to effectively assist picture knowledge extraction of distinct application domains. In brief, this research develops a picture knowledge extraction technology for the knowledge acquirers to efficiently and accurately acquire the picture knowledge and enhance reuse of pictures.