類神經網路影像檢索之研究---Web 2.0 影像及視訊檢索系統之建構(V)

Abstract

由於數位影像技術的提升與大容量儲存媒介價格的下降，大量的影像及視訊資料的儲 存愈來愈普遍。隨著網際網路的普及通訊品質的提升，有愈來愈多的數位媒體資料， 如文字、影像、視訊等被存放於網際網路上。如何精確、有效且快速地從網路獲得想 要的資訊，已是網路搜尋服務亟欲解決的問題。本研究首先針對Corel 影像資料庫， 於94 年度研發建立了以Visual keyword 作為影像索引的影像搜尋引擎。初步實驗結 果顯示，利用Visual keyword 搜尋Corel 影像資料庫的命中率超過九成。隨後，將此 技術拓展於網際網路上的影像搜尋上，建立了由20 台影像伺服器所組成的分散式影像 檢索系統，搜尋的範圍擴展到45,000 張以上的Internet 或WWW 影像。初步的搜尋命 中率已達六成以上，每次查詢平均需時6 秒。本年度擬研究在目前以圖像Visual keyword 為基礎的檢索（VBIR）技術上，結合「圖像標誌檢索」（TBIR，tag based image retrieval）技術的網際網路搜尋引擎，完成一個能搜尋網際網路影像與視訊的檢索系 統。應用「圖像標誌檢索」技術需將圖像先加注標誌，本計劃將研發Web 2.0 的標誌 系統：由進入網頁的使用者來協助注釋的工作，不斷地對小部份的影像加上標誌，再 以機器學習的方式將大部分的影像自動的加上標誌。由於網路上的使用者興趣與喜好 會隨時改變，因此人工方式的標誌將由各地上網使用者不斷的協助進行，隨時隨地適 應網際網路上多變化的環境。因此「Web 2.0 影像標誌系統之建構」將為本年度（97） 研究計劃書所規劃之創新研究工作。系統研發目標是建立由50 台伺服器所組成的分散 式搜尋系統，將搜尋的範圍擴展到200,000 張以上的Internet 或www 影像，以及檢索 速度提昇到一次影像詢問（image query）平均的需時在1 秒內。

The growth of the Internet and services has caused a corresponding explosion in the amount of media data that needs to be archived. The most important issue of web-based search is how to retrieve the desired media efficiently and correctly. In our previous related projects, we proposed a novel image index called Visual keyword and built an image search engine to retrieve the Corel gallery images. The experiment results show that the hit rate of the retrieved images is about 90%. Then, we built a web-based image search engine, which contains 20 PC based servers, to retrieve about 45,000 Internet images. The hit rate of the retrieved images is about 60%, and each query and retrieval needs 6 seconds in average. Since WWW is very dynamic, where users may come and go anytime, and users’ interests may change as time went by. The accuracy of image query varies largely without any rules or regularity. In this proposal, we plan to develop Web 2.0 based image tagging technologies to incorporate the current visual keyword based image retrieval prototype system for accurate and faster image search. Tag based image retrieval has been available for long time before the CBIR was proposed and developed. However, traditional manual tagging is a label and time consuming task. Thus, we plan to develop a Web 2.0 based system, so that Web browsers can assist the tagging and labeling of their favored images. Then, we will develop a machine learning scheme to propagate the “manual tagging” results to labeling more images on the Internet and/or WWW. When more and more images are tagged with proper labels, then more content based image retrieval work can be replaced by the tag based image retrieval work. Thus, the image retrieval speed can be greatly improved. On the system development part of this proposal, we plan to extend the current distributed image retrieval prototype system to a more realistic system, which will contains 50 PC based servers, and optimally reaches an average retrieval time to be less than one second.