Classifiying offensive sites based on image content

This paper proposes a method for helping to identify adult web sites by using the image-content as means of detecting erotic material. The image content is classified by investigating probable skinregions, and extracting their feature vectors. These feature vectors are based on color-, texture-, contour-, placementand relative sizeinformation for a given region. The importance of the different elements in the feature vector is determined by a genetic algorithm. For each picture, the algorithm gives the probability that a certain picture has erotic content. By mapping all the images in a web-site, and running the image-based classifier on the whole collection, we were able to set up a histogram of images with regards to the loglikelihood of erotic content for each image. Hence giving a good overview of the web-site’s content and at the same time leaving room for errors in the image-based classifier. The algorithm proved to be quite successful in our tests where all 20 sites where classified correctly. The image-based classifier is able to properly identify 89% of the evaluation images at an average processing speed of 11 images per second. Although this experiment focused on classifying adult web-sites, small alterations to the system can be done, enabling classification of other kinds of images and web-sites.