Intelligent Content-based Categorization of Web pages Using Combination of Textual, Structural, and Visual Features

Intelligent approaches have been applied recently in classification of Web pages, and as a specific application, in filtering of pornographic Web pages. The existing methods are mainly based on using textual features and in some cases visual features, but the main problem with these approaches is the high rate of over-blocking. In this paper, we have proposed a new intelligent model for filtering of immoral pages which employs three types of structural, textual, and visual features in a hierarchical structure of Bayesian and neural network classifier. As for structural and textual features, a list of keywords is extracted from the training pages and taking a correlation and statistical analysis over the features, we select the most effective features. As for visual features, we use skin color feature together with a set of part-based features extracted from the images within the page. The algorithm was tested over 1,295 Web pages including 700 porn pages (coming with text and image) in English and Persian language, and 505 non-porn pages including pages with medical, health, sports, etc. topics, and a 90% accuracy rate was obtained.