- 學位論文
以基因演算法進行影像檢索之推薦策略
A GA-based Recommender Strategy for Image Retrieval
摘要
隨著現代資訊科技與硬體設備的進步，網際網路的傳播與分享加乘效應，大量的數位多媒體資料，如音訊、影像、及視訊急遽的成長。過去數十年間，已有許多學者專家大力投入此關鍵技術的研究與開發。此一趨勢，引起我們研究的動機，期望提供一個能滿足多樣需求的影像檢索系統。本論文藉由探討文獻與實作成品中一些著名的內涵式影像檢索系統，分析檢索策略，尋找一些重要的主題；其中，有效的影像分割、快速的影像檢索模型及回饋機制的建立是核心目標，也成為我們在選擇影像特徵時最重要的評量依據。 本論文提出一個植基於區域內涵資訊的影像檢索系統，利用基因演算法進行影像檢索，並以回饋機制，提高檢索效率。本論文主要貢獻有：(一) 推導出一個有效的模糊推論影像區域分割方法。針對每一張影像，引入模糊推論法則，達成有效的影像區域分割，並以區域內涵高階的特徵向量代表該區域之特性。(二) 提出一個植基於基因演算法的影像檢索模型。在影像檢索過程，主要考量使用者視覺上感興趣的物件。依據使用者自不同影像中所選取的物件，以建立基因與染色體，並利用基因演算法進行影像檢索。(3) 提出一個影像檢索回饋探勘機制。系統會記錄使用者檢索過程，依據使用者檢索行為模式，將使用者所點選的區域，視為一筆交易，以進階分析使用者檢索影像的行為模式，據以提高未來影像檢索的效率。 本論文將詳述我們所提出的系統架構、方法及模型。實驗結果顯示，本論文提出的影像檢索模型，具有令人滿意的效果。同時，檢視目前結果與待深入探討之問題，本論文亦提出未來持續研究的方向與重點。
並列摘要
Along with the advanced information technologies, the availability of World Wide Web together with the rapid growth of photographic archives have attracted our research motivation in providing an efficient access to the digital image database through browsing and searching. Content-based image retrieval (CBIR) has been intensively studied in the last decades. In this dissertation, some existing CBIR systems and related literatures are reviewed, and the focusing issues of retrieval strategies are addressed. We emphasize the topics of effective image segmentation, fast image retrieval model and efficient relevance feedback mechanism. This dissertation proposes an efficient genetic algorithm-based image retrieval strategy that applies the regions of interest and relevance feedback mechanism to improve the retrieval efficiency. Three main contributions have been achieved. (1) A fuzzy inference model is presented to derive an effective image segmentation method. A set of higher order statistical descriptors are used to represent the characteristics of a region content. (2) A GA-based image retrieval model is proposed. To assist the users to formulate more precise queries, the proposed system allows users to choose specific regions from multiple images. According to the human preference, the combination of image content descriptors from the selected regions forms the chromosomes of the genetic algorithm used for retrieving the target images. (3) The user relevance feedback mechanism is employed to direct the advanced search. The selected regions by a user are transformed into a transaction record. Furthermore, the retrieval performance is further improved by mining association rules from the retrieval feedback. The system architecture and methodology are detailed in this dissertation and thorough experiments on different queries demonstrate the effectiveness and scalability of the proposed strategy. Meanwhile, the prospects of future research directions and topics are also given in the conclusion chapter.