Detection of region-duplication forgeries in digital images

Abstract

The introduction of digital cameras and digital image representation throughout these past few years have brought a revolutionary era to the world of digital photography. The use of film technology is by far obsolete. Digital image representation gave birth to a new problem in image representation; that of image forgery and manipulation. The production of powerful image manipulation software such as Adobe® Photoshop® made it much easier to manipulate photographs and the credibility of images has been brought to light. Image forensics has been in the lime light for a couple of years now. Due to the fact that image forgeries are so popular nowadays, the purpose of this dissertation is to implement a passive image forgery detection technique capable of identifying an authentic image from a forged image. Throughout the course of this dissertation, an intensive study about image forgery detection algorithms developed has been made, and upon these findings, a passive image forgery algorithm using a Weber Local Descriptor (WLD) implemented. The proposed system in this dissertation uses a Multi-Resolution WLD in order to extract image features from the chrominance components of a given image and uses these features in order to detect if this image is authentic or forged. This is made possible by using a dataset of images and extracting their image features using the designed algorithm, and with the use of these features train a Support Vector Machine which would in tum build a classifier capable of identifying patterns in both real and forged images. This classifier is then used in order to identify a test image as real or manipulated. A number of experiments have been conducted on the system and by the end of the dissertation a maximum percentage accuracy of 72.9% has been achieved. This accuracy achieved reflects the use of Multi-Resolution WLD algorithm rather than a single resolution algorithm and the test has been performed on the Cr chrominance channel.