Double JPEG compression detection and localization based on convolutional auto-encoder for image content removal

Abstract

Due to the widespread prevalence of JPEG images in daily life, their authenticity has often been investigated in detecting the traces caused by double JPEG (DJPEG) compression. This paper presents a novel method of DJPEG compression detection based on deep neural networks. After removing the image content using a convolutional auto-encoder (CAE), we separated the JPEG and DJPEG feature spaces into two aligned and non-aligned scenarios using a convolutional neural network (CNN) in the spatial domain. Besides detecting DJPEG compression for small-sized 64 × 64 image patches, the approach also lets us determine the manipulated regions. The proposed network was evaluated for different input images. Results on the standard raw images dataset (RAISE) demonstrated that the proposed methods outperform previous algorithms, especially in the non-aligned case. If an accurate determination of the first and second JPEG compression quality factors (QFs) is not possible, our algorithm is robust to their perturbations. Compared to one of the best existing methods, we achieved about 65% improvement in relative error reduction in some cases.