Impact of feature selection and kernel functions in classification for image acquisition forensics

• Main Author: Chua, Shaun Ming Jun.
• Other Authors: Sabu Emmanuel
• Format: Final Year Project
• Language: English
• Published: 2012
• Subjects: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
• Online Access: http://hdl.handle.net/10356/48486
• Institution: Nanyang Technological University

With the increased availability of digital image from the Internet and easy to get image editing software, the origin or the image source for these images have become a serious concern. The paramount value of the pictures lies in their ability to offer graphic credibility and convincing proof of facts in search of truth, which should be the ultimate goal of all litigation. To keep the integrity and ensure the accuracy and reliability of digital data, it is important to help establish a solid model on the characteristic of images obtained directly from its original source. This, in turn, will facilitate tampering forensics to determine if there has been any additional editing and processing applied to an image after it leave the source. Thus, investigate how different kernel functions used in SVM will affect the classification results in determining the origin of the image source. This project will focus on the dissimilarities in the image acquisition process of the imaging devices to develop two groups of features, namely color interpolation coefficients and noise features to obtain feature data from the image datasets that can jointly serve as good forensic features to help accurately trace the origin of the input images. Polynomial, Radial Basis Function and Sigmoid kernel functions were used in Support Vector Machine (SVM) learning based classifier to analyze the extracted features and recognize patterns with C-Support Vector Classifiers (C-SVC) for training and testing of data sets to identify acquisition device type. In the absence of any coding format knowledge other than the images itself, the analysis returned an average success rate of 99% in correctly detecting the original device type of the image file. Differentiating between images produced by standalone cameras, cell phone cameras, scanners and computer generated.