透過您的圖書館登入
IP:44.212.39.149
  • 學位論文

Robust Affinity Propagation for Picture Clustering

強健式親和性互動方法在圖像分群上的研究

指導教授 : 王家祥

摘要


近年來,人們使用數位相機拍照的情形越來越普遍,加上方便且便宜的特性,往往在不知不覺中拍下了許多數位相片。為了方便管理的目的,這些大量的相片應該要有良好的分類。但是,利用人工的方式比對且分類這些相片是非常困難、繁瑣的事情。基於上述的原因,如何利用計算機有效且自動的分群為數眾多及雜亂的圖片是一個很重要的研究挑戰。針對這個課題,在本篇論文中,我們提出了一個基於內容的圖像分群技術方法。我們的方法可以分成三個階段。第一階段中,我們使用SIFT feature以及MPEG-7 CLD feature來表示每一張要分類的圖片。SIFT feature可以有效地描述一張圖片的局部特徵,但是卻沒有考慮到顏色資訊,所以論文中我們使用MPEG-7 CLD feature來補足這個缺陷。第二階段中,我們利用了Affinity Propagation algorithm來當作主要的圖片分群演算法。此外,我們藉由加入了估計AP起始設定的步驟,而改善了AP的不穩定性。第三階段主要是針對第二階段的結果作後處理,我們將一些較小且相似的群合併起來。由本篇論文實驗結果顯示,對於為數1000張的圖片,我們的方法可以達到80%的ARI準確分數。當圖片數目擴大到3000張時,我們的方法仍然有70%的ARI準確分數。平均上而言,和不作任何修改的AP分群結果的ARI準確分數相比,我們所提出的方法增進了54%的幅度。總觀所有實驗結果,顯示出本篇論文提出的圖像分群方法是有效且強健的。

並列摘要


In recent years, taking pictures with digital camera becomes more and more popular. Unlike traditional photos, the cost of taking a digital photo is nearly free, so users often have a great number of digital pictures. For the purpose of management, those pictures should be well categorized. However, grouping a lot of pictures by hand is a difficult and boring task. As a result, how to use computer to automatically group numerous and chaotic digit pictures efficiently is an important research challenge. In this thesis, we propose a content-based picture clustering method for this topic. The proposed method can be separated into three phases. First, phase I extracts local SIFT (Scale Invariant Feature Transformation) features and global MPEG-7 CLD (Color Layout Descriptor) features from all input pictures. SIFT features can describe distinctive local characteristics of an image excluding the color information. Then, we add the color feature to compensate the problem of SIFT. In phase II, we adopt the Affinity Propagation (AP) algorithm as our image clustering method. Further, we improve the instability by appending an estimating step that can evaluate a more suitable initial setting of AP. Finally, phase III is the post-processing stage that merges those small and similar groups produced in phase II. The experimental results show that the proposed method has over 80% ARI accuracy score for 1000 pictures. When the size of the dataset expands to 3000 pictures, the ARI accuracy score is still 70%. On average, our proposed method has 54% improvement in terms of ARI score, as compared to the pure AP algorithm. All the experimental results show that the proposed clustering algorithm is effective and robust.

參考文獻


[1]. Y. Jing and S. Baluja, “VisualRank: Applying PageRank to Large-Scale Image Search,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 30, no. 11, pp. 1877-1890, 2008.
[4]. D. Cai, X. He, Z. Li, W.Y. Ma, and J.R. Wen, “Hierarchical clustering of WWW image search results using visual, textual and link information,” in Proc. ACM International Conference on Multimedia, pp. 952-959, 2004.
[5]. T. Liu, C. Rosenberg, and H. A. Rowley, “Clustering Billions of Images with Large Scale Nearest Neighbor Search,” in Proc. IEEE Workshop on Applications of Computer Vision, pp. 28, 2007.
[6]. B. J. Frey and D. Dueck, “Clustering by passing messages between data points,” Science, Vol. 315, No. 5814, pp. 1136800-976, 2007.
[7]. B.J. Frey and D. Dueck, “Non-metric affinity propagation for unsupervised image categorization,” in Proc. International Conference on Computer Vision, pp. 1-8, 2007.

延伸閱讀