Mining and Classifying Images from an Advertisement Image Remover

Abstract

AdEater is an early browsing assistant that automatically removes advertisement images from internet pages. It works by generating rules from training data and implementing these rules when browsing the internet. Advertisement images on web pages are replaced by transparent images that display on the image the word “ad”, and where images are misclassified, non-advertisement images on a webpage will also be replaced by transparent images displaying “ad”. This paper critically examines the dataset derived from a trial of AdEater and tries to build a robust image classifier. We apply data mining techniques to uncover associations between features of advertisements and non-advertisements and try to predict whether the images are advertisements or non-advertisements based on three classification methods. We achieve classification accuracy of 96.5%, using k-fold cross validation to train and test the model.