點對點網路上以內容為基礎的多媒體資料索引及搜尋技術

Abstract

點對點網路(Peer-to-Peer, or P2P)由於其低成本及簡單具延展性的特性，已經越來越 被廣為應用在資料分享上。然而，大多數的系統目前只支援基於檔名或是關鍵字的查 詢，使用者不能根據檔案的內容或提供更具語意的條件來進行搜尋，亦即使用者必需很 明確的知道所欲尋找的資料當中的關鍵字，否則將查詢不到。以內容為基礎的查詢則能 提供較友善(具語意)、彈性且複雜的多媒體資料查詢。由於P2P系統是一個動態的、分散 的系統，在P2P環境下進行複雜且具語意查詢，是十分具有挑戰性的。 點對點的檔案系統依架構來分，可以分成無結構式系統(Unstructured P2P )以及結構 式系統(Structured P2P)等方式。無結構式P2P沒有中央管理的索引服務，它採用了 Flooding方式和Random Walker機制，此方法將造成網路流量急劇增加。結構式系統主要 是透過雜湊函數(DHT)來算出儲存資源的節點，而查詢某資源時，也採用同樣的雜湊函 數就能找到目標節點，因此搜尋的準確性得可以保證，具有良好的可擴展性。結構式P2P 雖然具有上述優點，但DHT演算法只適合於準確的查詢，很難支援多關鍵字查詢、範圍 查詢、甚至具語意等複雜查詢。 此外，訊息通報服務(Event Notification Service) 也是近年來漸受重視之新技術之 一，如何讓資訊的發怖能自動傳遞、撮合及通知有興趣的使用者，對以服務為導向的應 用來說更為重要。在P2P 網路上由於每個使用者也都是內容提供者，困難度更高。其 中的關鍵技術包括：發佈/訂閱系統的索引、比對(Match)演算法、路由(routing)演算法等。 在此計劃中，我們預計以三年為期完成以下三個目標： 1. 以內容為基礎的多媒體 資料查詢 2.以使用者訂閱為導向的多媒體資料發佈通知技術，3.高低階特徵值整合，並 加入以使用者偏好為導向的自動通知技術。第一年主要是研究如何運用各種有效且能快 速取得的多媒體特徵以及如何索引的技術，使得發展之技術能適用於點對點的多媒體網 路上。第二年我們將加入使用者訂閱的功能，針對使用者訂閱的內容及新發佈的多媒體 資料內容，研發如何分散地管理以及建立相互關連的索引，使得在點對點網路上有任何 新資料發佈時，能自動通知符合訂閱要求的使用者。第三年我們將研究如何根據前兩年 所發展的技術，整合以內容為基礎的多媒體資料查詢與以高階語意描述的使用者訂閱功 能，完成多媒體內容自動傳至使用者的技術，及相關應用系統的建構。在建構完成的系 統中，我們擬加入以使用者偏好為導向的多媒體資料發佈自動通知功能。

Peer-to-peer (P2P) networks have gained popularity as a simple, low-overhead means for the dissemination and sharing of information. However, the retrieval facilities of most existing P2P systems are limited to queries based on a unique identifier or a small set of keywords, which means that the end user cannot retrieve the content unless he knows its unique name or specific keywords. Content-based multimedia retrieval (CBMR) system which supports more flexibility and complex query types has attracted a great deal of research attention. However, most current CBMR systems are based on the centralized computing environments. Supporting CBMR in the P2P networks is still a challenging problem. Traditional keyword based query routing uses flooding based on unstructured P2P systems, such as Gnutella, or uses distributed hash table (DHT) based on structured P2P systems, such as Chord, CAN, and Pastry. While these approaches provide accuracy of query results, they are not applicable for content based retrieval due to large network traffic required, limited to exact-matched queries, high overhead in reorganization for node leaving/ joining, or lack of flexibility and scalability, etc. Besides, Event Notification Service which automatically notifies users for new published content has become important for service-oriented applications. It is even more challenging in P2P networks for which every one can be content providers. The related research topics include publish/subscription indexing, match, routing algorithms, etc. In this project, we plan to design a P2P-based video processing and indexing framework that supports content-based video retrieval. There are three goals that we would like to achieve in three years: 1. Content-based multimedia information retrieval. 2. User-subscription oriented notification for multimedia information publishing. 3. An integrated framework with the adding of user-preference oriented auto-notification technology. In the first year, we will focus on the study of various video feature extraction and indexing technologies that are suitable for the P2P networks. In the second year, we will study user query/subscription feature extraction and distributed management technologies. We will also investigate various indexing and routing mechanisms to support automatic notification with minimum network traffic requirement. In the third year, our study is focused on proposing an integrated framework that will bridge the gaps between low-level video features and semantic-level user queries/subscriptions. Based on the framework, user-preference oriented auto-notification technology will explored by analyzing user query context. We will implement a system for the integrated framework, and will conduct some experiments to demonstrate our framework in a large-scaled P2P network.