Detecting Doctored Region in JPEG Image using Convolutional Neural Networks

Abstract

Many digital pictures are used as evidence in criminal investigation. It is very important to check whether they are doctored or not. This paper proposes a method for detecting doctored region in JPEG image using a convolutional neural network (CNN). In the proposed method, DCT coefficients are input to the CNN. Its output is a binary segmented image in which doctored and non-doctored regions are shown with white and black pixels, respectively. In our experiment, 45 types of CNN models were created and compared. The detection accuracy of the best model achieved 0.63 in terms of F-measure, which is larger about 2.3 times than that of our preliminary method based on support vector machine (SVM).