Detecção automática de conteúdo ofensivo na web

Abstract

The World Wide Web is a huge source of diverse information, including offensive material such as pornography related content. This poses the problem of automatcally detecting offensive content as a way to avoid unauthorised access, for instance, by children or by employees during working hours. Although this sort of information is published in many forms, including text, sound and video, images are the most common form of publication of offensive content on the Web. Detecting offensive images can be considered as a classification problem. Given that images are part of Web pages, textual information can be used as important evidence along with the content extracted from images, such as colour, texture and shapes. This dissertation proposes two distinct approaches for automatic detection of offensive images on the Web. The first is based on image content, specifically colour. The second approach is based on textual terms extracted from the Web page that present the images. After evidence extraction the classification is performed using the SVM technique, based on a collection of 1000 offensive images and 1000 non-offensive images for training. Experiments carried out have shown that both approaches are effective, although they rely on simple algorithms for extracting evidences related to the images.