應用於拍賣網頁分類目錄下的影像搜尋系統

近年來，隨著多媒體資訊的進步，數位相機、具有照像功能的手機等相關科技產品的普及化，許多研究學者開始採用影像的內容來做為圖片搜尋的依據。本論文主要是將此方法應用在Yahoo拍賣網頁上，並且在特定的分類目錄下，實做一個影像搜尋系統。本搜尋系統，提供使用者上傳或點選資料庫中影像做搜尋。傳統影像搜尋的方法是使用整張圖的影像特徵，而本論文使用了二種物件偵測演算法進行實驗，一種是利用邊緣偵測找出物件邊緣，再利用Principle Component Analysis(PCA)計算邊緣點聚集的區域，稱為Edge+PCA，此區域即可能是影像中重要的物件區域；另一種則是經過AdaBoost演算法，從大量簡單特徵為主的弱分類器中合併有效的弱分類器形成一個快速並精確的強分類器，再將多個強分類器組成Cascade結構，稱為AdaBoost偵測器。本論文利用Yahoo拍賣網頁抓取的資料以驗證本系統的可行性，實驗證實Edge+PCA偵測的方法和AdaBoost偵測都有70%的準確率，並且利用物件區域的特徵進行搜尋，在搜尋的R-Precision較利用整張影像特徵搜尋有顯著的改進。本論文亦實作了文字搜尋與文字及影像合併搜尋，與上述影像搜尋比較其效能。

關鍵字

影像內容搜尋系統；色彩直方圖；物件偵測； PCA ； AdaBoost

並列摘要

The purpose of this paper is to apply content-based image retrieval to online auction system with build-in directories. The development of an intelligent category-and content-based search engine is presented, in which it allows a user to upload an image file or choose one image from the database to search for visually similar auction items. Traditional content-based image search methods use features extracted from the whole image. These features contain many noises from the background and result in poor performance. We propose two novel approaches to extract features from the most likely object region in an image. The first approach, called ”Edge + PCA (Principle Component Analysis)”, assumes that there is rich edge information around the object in an image. A representative region is calculated from the edge map of an image using PCA. The second approach is to apply AdaBoost algorithm to design a specific object detector for each category. In the experiments, images and related webpages are crawled from Yahoo online auction. Both object region detection methods have about 70% accuracy and the search precisions using object features have obvious improvement over those using features of the whole image. In addition, we combine image search with text search and examine its performance against image-only and text-only search.

並列關鍵字

Content-based image retrieval ； Color histogram ； Object detection ； PCA ； AdaBoost

國際替代計量

應用於拍賣網頁分類目錄下的影像搜尋系統

全文下載

主題瀏覽