Abstract
隨著多媒體資料庫, 視訊資料庫以及隨選視訊之愈來愈流行, 視訊資料的擷取成為一個相當重要的課題. 與傳統的關鍵字或數字的比對查詢, 以視訊本身的內容與內涵來查詢視訊資料是一種更有效好用的資料擷取方式.透過視訊中內涵物件的空間與時間關係的描述, 符合這些描述的影片內容, 就可以經由查詢的處理, 提供給查詢者當做答案.在這篇論文中, 主要對於內涵式的查詢進行研究. 我們將它分為三個課題: (1) 視訊索引:內涵式視訊擷取, 查詢是基於視訊內之特徵進行. 這些特徵必須被萃取出來, 並索引化以供查詢處理之用.(2) 視訊擷取: 基於這些特徵, 我們需要設計視訊擷取的方式, 如查詢語言, 視訊瀏覽系統或其他介面. (3) 查詢處理: 根據所設計的索引與視訊擷取方式, 提供查詢處理的方法.最重要的問題是這些特徵萃取的方法. 第一, 將視訊根據鏡景來分段是建立視訊索引的基本工作. 我們提出了一個新的方法來偵測鏡景的變換. 此方法是基於 MPEG 壓縮視訊的格式中隱含的資訊來進行. 好處是可以避免逐一畫面耗時的比對. 並且不需將視訊資料解壓縮. 因此, 比對速度很快.視訊擷取方面, 我們據視訊內涵的特性提出了一個內涵式視訊查詢語言CVQL.在 CVQL 中, 使用者可以描述視訊內容物件的時間空間關係進行查詢. 我們定義了一組視訊函數來描述這些關係. 我們同時也設計了一個圖形化查詢介面來簡化查詢的描述. 查詢的處理是以一個過濾式的前置處理與行為式條件比對來加速處理的速度. 我們提供了一個相關於此查詢處理的索引, 來達成這個功能.The retrievals of video data become an important issue due tothe widerequirements from the multimedia databases, videodatabases, andvideo-on-demand systems. Comparing to traditionalkeyword-based dataretrieval, content-based retrieval provides afruitful way to retrievevideo data. By specifying the temporaland spatial relationships ofcontent objects, the predicatedescribes video contents and then thevideos satisfy thepredicate are retrieved as query results. In this proposal, westudy the content-based retrieval for videodatabases. There arethree significant issues. (1) Video indexing: Forcontent-basedretrieval, query is based on the features of video contents.Thefeatures are extracted and indexed for the query processing. (2)Videoretrievals: Based on these features, the approach for videoretrievals isnecessary, such as query languages, browsingsystems, and other interfaces.(3) Query processing: Evaluatingthe query by accessing the video indexes. The most importantissue is the extraction of the features from videocontents.First, video segmentation is an elementary work for videoindexconstruction. We propose a new approach to detect shotchanges for videosegmentation, which is based on the processingof MPEG compressed videodata. This approach takes advantages ofthe information implied in thecompressed data. The similaritymatching on raw video data can be avoided.Therefore, theprocessing is efficient. For video retrievals, we analyze thefeatures of video contents andpropose a content-based videoquery language CVQL. In CVQL, spatial andtemporal relationshipsof content objects are used for the specificationof querypredicates. A set of operations for specifying temporalandspatial relationships in queries is defined. By theseoperations,characteristics of video data can be used for queryqualification. Userscan express the semantics of demanded videodata by combining the proposedtemporal and spatial operations.We also implement a graphical queryinterface (GUI) whichsimplifies the use of CVQL in the prototype videodatabasesystem. To provide an easy way to retrieve video data, asimple browsingsystem is implemented. Users may query by animage example. The queryimage is matched with the key framesextracted from the video shots. Theshots of the matched keyframes are retrieved for users.For the processing of CVQL, anelimination-based approach is proposed. The query processingeliminates the unqualified videos and frames bychecking theexistence of the content objects. It reduces the highcostprocessing of the evaluations of video functions video byvideo and frameby frame. Moreover, approximate querycapabilities are provided byrelaxing the strict matching oncontent objects and video functions. Inthis case, the queryresults are shown to the user with the value whichrepresent thematching degree of the predicate. To improve the evaluation ofvideo function, we further propose anapproach to avoid frame byframe evaluation. In this approach, we firstanalyze thebehaviors of content objects by their spatial andtemporalrelationships. The behaviors are then classified intoseveral basic types. Aset of indexes is constructed for them toaccelerate the evaluation of videofunctions.