September, 2010 Student : Ya-Fen Cheng |
|||
【研究簡介】 中文摘要: 隨著科技不斷的發展、電腦與網路應用的快速普及化,眾多問題的解答,皆可利用強而有力的文字搜尋引擎,目前對於以文字進行檢索資料,已可達到精確的程度,例如: Google搜尋引擎。但隨著具照相功能的手機及數位相機等相關產品的普及化,存在於網頁或其他媒體上的龐大影像資訊不斷的增加,文字已無法詮釋某些影像的概念和意境,因此「以圖找圖」的影像搜尋系統因應而生。本論文主要研究為圖像商標檢索。在當今全球化經濟中智慧財產權 (Intellectual Property Right, IPR) 已日益重要,為了保護商標所有人的權益,搜尋 (Search) 和監看 (Watch) 是否有仿冒或盜用圖像商標的情況具有重要的應用價值。我們利用三種影像局部特徵子作為特徵擷取與比對的方式,分別為 SIFT、PCA-SIFT 和 SURF 特徵,主要目的要找出對於圖像商標有最好效能的特徵子。商標資料庫中有著相當大量的影像,而影像中所包含的局部特徵子亦非常多,因此利用窮舉法搜尋相當費時,為了加快搜尋的速度,我們利用了字彙樹 (Vocabulary tree) 作為影像搜尋的方法,但是當速度變快時,相對的準確率也會降低,因此要在準確率與速度之間取得一個平衡。本論文分別利用改良的貪婪的N條最佳路徑搜索 (Greedy N-Best Paths Search) 和幾何校正 (Geometric Rectification) 提升準確率。在 200 張查詢影像以及台灣經濟部智慧財產局 27,610 張資料庫影像的實驗中,準確率可高達九成左右。Abstract : Attributing to the development of informational diversification, users can directly acquire a large number of information which is useful and surfed immediately from the browser. Nowadays, the technique of text retrieval has been developed more maturely, such as “Google search”. However, text-only search is not enough for the variety of resource from the internet, because there are more and more types of information in the world, such as video, audio and the combinations of them. Nowadays, intellectual property right in the globalization economy has received much attention, and watch for infringed trademarks is one of the most important issues. This thesis focuses on figurative trademark search on the image database collected at the Ministry of Economic Affairs Intellectual Property Bureau, Taiwan. To search for similar trademarks using image content, we make use of three kinds of features including SIFT, PCA-SIFT and SURF for matching between the images. The first goal is to find out which feature is most suitable for the trademark database. Because of the large amount of features extracted from the images, the off-line training and on-line processing both spend a lot of time to finish. To improve the performance, we use the vocabulary tree for reducing the search time. While the search efficiency is improved, the search accuracy becomes slight poor. For this purpose, two modifications of the greedy N-best paths search and geometric rectification are used for reclaiming the accuracy. In the experiments, 200 images are used as queries for searching 27,610 images in database, and the accuracy of the performance can be up to 90% or so. |
|||
【相關流程】 |
|||
![]() |
|
||
【結論】 本論文對 27,610 張資料庫影像以及 200 張查詢影像各別使用SIFT 特徵、PCA-SIFT 特徵和 SURF 特徵,此三種特徵擷取方法進行比較,由上述實驗可明顯得知,在我們的資料庫中使用 SIFT 特徵和 PCA-SIFT 特徵於窮舉法的實驗以及利用 Count Match 相似度比對方法時,兩者有著不相上下的準確率,但是SIFT 特徵仍然略勝一籌;然而,SURF 特徵則比前兩者影像特徵略於遜色。由此可見,SIFT 特徵在我們的資料庫中(圖像商標)確實比其他兩種特徵更具強健性。 由於,在窮舉法的實驗中會花費過多的時間,於是我們使用了字彙樹,加速搜尋所需的時間。一旦速度增快之後,準確率必然會受到影響而下降,對於 PCA-SIFT 特徵以及 SURF 特徵而言,此量化的動作無論是在 Binary feature vector、L1-norm 或 L2-norm 等相似度比對方法時,皆對他們的準確率造成了大幅度的影響;而對於SIFT 特徵而言,在 Binary feature vector 相似度比對方法中,只有小幅度的下降。 因此必須在準確率與速度之間取折衷,於是我們修改了貪婪的 N 條最佳路徑搜索以及利用幾何校正使影像重新排序,希望藉以挽回因增快搜尋速度而小幅度下降的效能,由上述實驗證明此兩種方法皆可成功的提升失去的準確率。 |
|||
【論文全文】 |
|||
Figurative Trademark Retrieval System based on Local Descriptors【PDF】 |
|||
Figurative Trademark Retrieval System based on Local Descriptors【word】 |
|||
【口試投影片】 |
|||
Figurative Trademark Retrieval System based on Local Descriptors【ppt】 |
|||
【主要相關論文】 | |||
● D. Nistèr and H. Stewènius, “Scalable Recognition with a Vocabulary Tree,” Computer Vision and Pattern Recognition, Volume 2, pages 2161-2168, 2006. ● G. Schindler, M. Brown, and R. Szeliski, “City-Scale Location Recognition,” In Conference on Computer Vision and Pattern Recognition, pp. 1-7, June 2007.【PDF】 |
|||
Advisor : jcliu@ncnu.edu.tw Jen-Chang Liu
VIP Lab. @ CSIE NCNU |