Local DCT-Based Deep Learning Architecture for Image Forgery Detection

Abstract

A convolutional neural network model efficient in forgery detection in images regardless of the type of forgery is proposed. The AC coefficients in the block DCT of the entire image are analysed for the suspected forgery operation. The feature vector is extracted from the non-overlapping blocks of size $8$ $\,\times \,$ $8$ of the image. It consists of the standard deviation and non-zero counts of the block DCT coefficients of the image and its cropped version. The image is first converted to YCbCr colour space. The feature vector is extracted for all three channels. We then supply this feature vector as input to the deep neural network for detection. We have trained the DNN using CASIAv1 and CASIAv2 datasets separately and tested them. The train test ratio used is 80:20 for experimentation. Experimentation results on standard datasets, namely CASIA v1 and CASIA v2, reveal the efficiency of the proposed approach. A comparison with some of the existing approaches shows the proposed approach’s performance in terms of detection accuracy.