Transcript 下載/瀏覽
Automatic Website Summarization by Image Content: A Case Study with Logo and Trademark Images Evdoxios Baratis, Euripides G.M. Petrakis, Member, IEEE, and Evangelos Milios, Senior Member, IEEE IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 20, NO. 9, SEPTEMBER 2008 Date : 2009/10/29 Speaker : Chin-Yen Yang 南台科技大學 資訊工程系 Outline 2 1 INTRODUCTION 2 IMAGE FEATURE EXTRACTION 3 PROPOSED METHOD 4 EXPERIMENTAL RESULTS 5 CONCLUSIONS 1. INTRODUCTION We introduce the concept of image-based summarization A fully automated image-based summarization approach is proposed The evaluation of the method on corporate Websites is presented 3 1. INTRODUCTION (C.) Logos and trademarks are important characteristic signs of corporate Websites A recent contribution reports that logos and trademarks comprise 32.6 percent of the total number of images on the Web 4 2. IMAGE FEATURE EXTRACTION Intensity histogram Radial histogram Angle histogram 5 2. IMAGE FEATURE EXTRACTION (C.) 2.1 Image Representation 6 3 PROPOSED METHOD 7 3 PROPOSED METHOD (C.) 3.1 Image Information Extraction 1. Link information MaxDepth 1 LinkDepth Depth MaxDepth 2. Text Information This information is displayed together with images or can be used for searching the Web 8 3 PROPOSED METHOD (C.) 3.2 Logo and Trademark Detection Training the decision tree using histogram features outperforms training using raw histograms 9 3 PROPOSED METHOD (C.) Similarity detection Three attributes corresponding to three histogram intersections, and one attribute corresponding to the euclidean distance of their vectors of moment invariants The decision tree was pruned with a confidence value of 0.1 and achieved a 93.89 percent average classification accuracy 10 3 PROPOSED METHOD (C.) Image clustering 3.3 Duplicate Logo and Trademark Detection From each cluster, one image is selected to represent the cluster in the summary 11 3 PROPOSED METHOD (C.) 3.4 Logo and Trademark Ranking Probability Instances Depth Image Importance = Probability*Depth*Instances 12 3 PROPOSED METHOD (C.) 3.5 Image-Based Summarization Cluster Importance = Image Importancei image.icluster 13 4 EXPERIMENTAL RESULTS 14 4 EXPERIMENTAL RESULTS (C.) 15 5 CONCLUSIONS First by extracting images with high probability of being logos or trademarks Clustering similar images together and by ranking images in each cluster by importance The most important image from each cluster is included in the summary 16 5 CONCLUSIONS(C.) 76 percent detection accuracy 85 percent classification accuracy 64 percent summarization accuracy Future work includes experimentation with larger training data sets and image types for improving the performance machine learning 17 南台科技大學 資訊工程系