疗铺桥惹铺肆.PDFVIP

INDEPENDENT COMPONENT ANALYSIS FOR UNDERSTANDING MULTIMEDIA CONTENT Thomas Kolenda, Lars Kai Hansen, Jan Larsen and Ole Winther Informatics and Mathematical Mo deling, Building 321 Technical University of Denmark, DK-2800 Kongens Lyngby, Denmark Phone: +45 4525 3920,3923,3889,3895 Fax: +45 4587 2599 E-mail: thko,lkh,jl,owi@imm.dtu.dk Web: eivind.imm.dtu.dk Abstract. This pap er fo cuses on using indep endent comp onent analysis of combined text and image data from web pages. This has p otential for search and retrieval applications in order to retrieve more meaningful and context dep endent content. It is demon- strated that using ICA on combined text and image features pro- vides a synergistic eect, i.e., the retrieval classication rates in- crease if based on multimedia comp onents relative to single media analysis. For this purp ose a simple probabilistic sup ervised clas- sier which works from unsup ervised ICA features is invoked. In addition, we demonstrate the use of the suggested framework for automatic annotation of descriptive key words to images. INTRODUCTION Understanding the structure of multimedia data is increasingly imp ortant for retrieval, indexing and search. With the advent of the advanced MPEG standards [25] content and context sensitive to ols will b ecome indisp ensable comp onents of the webminers multimedia to olb ox. Content based image retrieval is a highly challenging asp ect of multime- dia analysis [9]. The task is hard b ecause of the limited understanding of the relations b etween basic image features and abstra