Automated Image Forensics Based on Deep Learning for Discriminating Photorealistic Computer Graphic and Photographic Images

Abstract

The rapid advancement in the field of digital image forensics from the recent few years is becoming very crucial nowadays. As it become easier to generate computer graphic images and forge the whole of the image or just part of image to perform illegal activities. Distinguishing Computer generated stuff with the Natural images is quite difficult task with naked human eye.

In this research thesis we proposed the CNN-based Neural Network model for the identification of images. For the classification of images Columbia image dataset is used which includes Photorealistic Computer Generated (PRCG) images and Photographic Images (PIs). Original, resized, filtered and patched images with different modification are used to feed in Convolutional Neural Network (CNN) before training and testing phase.

The proposed method used five convolutional layers. All of the experimental results and analysis shows that proposed method achieved 99% training accuracy and 98.5% validation accuracy on different types of images which is although sufficient enough for this proposed method with adequate amount of dataset.