convolutional autoencoder

With the increasing interest in Joint Photographic Experts Group (JPEG) image compression, one of the most important issues in digital image manipulation is finding a proper method to detect double JPEG compression. This paper introduces a trained adaptive filter based on spatial-domain convolutional autoencoder (CAE). This filter can remove interference information caused by image content to have a more accurate detection. The convolutional neural network (CNN) has been widely employed for accurate image classification; therefore, a CNN is used in the classification part of the proposed algorithm. The proposed model is based on consecutive CAE with CNN, which is able to provide acceptable detection accuracy and sensitivity to quality factors (QFs) in two scenarios, i.e. aligned and non-aligned forgeries. This model improves the sensitivity to quality factors by up to 86% in the relative error reduction (RER) rate in some cases. Other experiments such as manipulation localization on the RAISE dataset have been performed to evaluate the proposed method. These results show the superior performance of this method compared to similar algorithms in the situations that the quality factor of the second compression is greater the quality factor of the first compression.