摘要
以內容為基礎之影像查詢(CBIR)的研究內容包含特徵選取、特徵表示以及結果比對。常被取用的特徵表示包含形狀、顏色、空間位置、及紋理等。一個好的影像表示法必須能夠克服影像的位移、旋轉、以及放大或縮小等問題。甚至於對影像在一定程度內的損毀下也必須能夠有好的比對結果。在以影像內容為查詢基礎時這些要求均為相當重要的議題。 考慮以上的要求，本論文提出三個以影像內容為基礎之查詢系統。首先是針對物體邊緣的研究，我們使用一快速的邊緣點偵測演算法來偵測出影像中所有可能的邊緣點，並提出一新的物件表示法—爬山式序列表示法(Mountain Climbing Sequence (MCS))。此表示法對於前面所提之影像中的位移、旋轉、以及放大或縮小等問題都可以達到不變 性的要求。另外，由於邊緣點的偵測就目前的研究經驗上並無法保証能夠找一物件的完整外形，因此我們也將嘗試在現有的外形特徵表示法下，克服物件外形不完整抽取的情況，甚至於在物件少部份被遮蔽的狀況下我們所提的系統也能得到良好的比對結果。其次是邊緣點與相鄰邊緣點關係的研究，在真實影像中，我們很難精確的偵測出影像中物件的完整的邊緣點，因此，我們計算影像中邊緣點與邊緣點間之關係，建立一方向-距離之長條圖，用以表示此真實影像，此表示法能克服查詢時所須面對的縮放及旋轉問題。 最後，我們結合了許多各類不同之影像特徵來表示一真實影像，包含方向—距離長條圖、區域邊緣點關係、小波係數、顏色分佈長條圖、顏色飽合度分長條圖及亮度分佈長條圖共1仟多維度的特徵。在此高維度的特徵比對過程，我們利用一個修改後的AdaBoost演算法，快速且有效率的查詢出相似影像，實驗證明，我們所提出的方法確實能達到在查詢影像時所須的條件—快速及有效率。
並列摘要
Because of recent advances in computer technology and the revolution in the way information is processed, increasing interest has been developed in automatic information retrieval from huge databases. In particular, content-based image retrieval (CBIR) has become a hot research topic and, consequently, improving the technology for content-based querying systems is of more challenge. The work of content-based image retrieval includes selection, object representation, and matching. A good image representation should meet some requirements, including invariance to translation, rotation, scaling and reversion, and sustaining deformation of query images. In this dissertation, we proposed three robust and efficient methods for CBIR. First, an efficient and robust shape-based image retrieval system is proposed. We introduce a shape representation method, the Mountain-Climbing-Sequence (MCS), which can be used to make the retrieval to be invariant to translation, rotation, scaling, and reversion. Second, a new feature called orientation-distance histogram is introduced, and a CBIR system based on this feature is proposed. This system transforms the RGB color model of a given image to the HSV color model and detects edge points by using the H-components, and then evaluates the orientation-distance histogram for each of the detected edge points to form a feature vector. With normalization of feature vectors and the use of MCS sequences, this system is of invariance to scaling and rotation. Last, we collect a variety of features for representing real images, including orientation-distance histogram, local edge map, wavelet coefficients, color information, and intensity information, forming feature vectors of dimension up to one thousand. Over such a large set of features, we use an improved version of the AdaBoost algorithm to select the most important features indicated by the user, and so as to efficiently achieve effective retrieval results.
參考文獻
延伸閱讀
- 吳靜宜、林麗娟、曾元顯（2004）。影像內容檢索需求與使用之研究。圖書資訊學刊，1(4)，21-35。https://www.airitilibrary.com/Article/Detail?DocID=16067509-200403-1-4-21-35-a
- 林頌堅（2008）。以內容特徵為基礎的網路影像檢索系統及相關技術的探討。臺北市立圖書館館訊，25(4)，12-33。https://www.airitilibrary.com/Article/Detail?DocID=10112081-200806-201312160034-201312160034-12-33
- Chang, L. Y. (2009). 以內容為主的監視影像搜尋系統 [master's thesis, National Taiwan University]. Airiti Library. https://doi.org/10.6342/NTU.2009.00630
- Li, C. Y. (2006). 利用區域資訊進行內涵式影像檢索的檢索歷程學習機制之研究 [doctoral dissertation, National Tsing Hua University]. Airiti Library. https://www.airitilibrary.com/Article/Detail?DocID=U0016-1303200709470263
- Lin, M. F. (2011). Strategy Improvement for Content-based Image Retrieval [doctoral dissertation, Chaoyang University of Technology]. Airiti Library. https://www.airitilibrary.com/Article/Detail?DocID=U0078-1511201110381237