A Novel Framework for Web Pages Classification
Ruiguang Hu, Weiming Hu · Advances in intelligent systems research/Advances in Intelligent Systems Research · 2013
Regarding web pages classification, lots of algorithms have been proposed and promising classification performances have been achieved [1][2][3][4][5].Unfortunately, to and Weiming Hu 2 Abstr act.In this paper, we propose a novel framework for classifying web pages containing images and text.Valid images are first chosen by the FOrward Com-pArison of Relative Sizes Sorting(FOCARSS) algorithm, and each valid image is represented by the mid-level feature vector generated by the Bag-Of-Features model.Taking these feature vectors of valid images in a web page as instances of a bag, Multi-Instance Learning is utilized to conduct the image-based web pages classification.Regarding the text information, Bag-Of-Words model is used to conduct the text-based web pages classification.Subsequently, score-level fusion schemes are used to fuse these two kinds of heterogeneous information.Experimental results on a representative dataset demonstrate that our framework can definitely take full advantage of image and text information and improve final classification performances. Keywor ds: Score fusion • Multi-instance learning • Bag-of-features 1 Intr oductionWith the rapid development of the Internet and the extensive use of intelligent devices, such as smart phones and cameras, massive amounts of images have been emerging on the web, consequently, the content of web pages are extremely multitudinous.Additionally, web page designers prefer to utilize images to express the theme of a page.In extreme cases, a web page contains lots of images while only a few or even no text, and this kind of web pages take a increasing proportion, which can be found easily.