Image Forgery Detection and Classification using HMM and SVM Classifier

Abstract

In the modern era of digital publishing and imaging, images are retouched and manipulated to increase and enhance the aesthetic beauty of the images. Several images are also morphed before publishing to incorporate extra information in the image. Scene morphing, face morphing are other two significant retouching. Therefore finding the authenticity of the images is largely difficult. The problem becomes quite critical due to various sources of image capture. There are different image sensors with different resolutions and encoding. There is no fixed standard for the same. Many images are also compressed or resized before publishing. Therefore identifying the forgery in the images is not only challenging but also at the same time quite difficult. In this paper we present a novel approach for image forgery detection. We observe that a non morphed and non forged image shows homogeneity in non spectral domain. This homogeneity is lost when any forgery or morphing is applied on the images. We therefore apply a set of transform over the images. We combine DCT statistics, LBP features with curvelet statistics and Gabor transform of the images to represent an image in the transformed domain. CASIA image dataset with seven thousand authentic and same numbers of tempered images is used to verify the technique. We divide the dataset into equal halves to perform training and testing. Transformed images are used to train Hidden Markov model as HMM can extract probabilistic state information from a large statistical model. A test images is tested in transformed domain by HMM with log likelihood estimator. In case HMM returns an indeterminist result or multiple subset of result, the transformed test image is tested with two class SVM classifier with RBF kernel. Results show that the accuracy of the system is over 95% for 3500 test instances.