Multi-scale Deep Network Based Double JPEG Compression Forensics

With the popularization of image processing tools and the rapid development of multimedia technology,digital images have been widely used in various fields as the main information carrier.Due to the diversity of image processing tools,people can easily modify an image in any desired way.As a result,the current digital technology has seriously affected the credibility of visual imagery in many fields such as journalism,military,justice,commerce,medical applications,and academic research.Due to favorable compression performance and reconstruction quality,JPEG has become the most widely used image format.Therefore,in the field of image blind forensics,the importance of tamper detection for JPEG images is self-evident.This article has carried on the thorough research to the method of double JPEG compression forensics and has proposed two kinds of multi-scale deep network based JPEG image forensic algorithms.In the field of JPEG image forensics,extracting effective statistical characteristics from a JPEG image for classification still remains a challenge.Effective features are designed manually in traditional methods,suggesting that extensive labor-consuming research and derivation is required,and the classification effect of these features is often limited.In this paper,a multi-scale deep discriminative network based algorithm for JPEG image forensics is proposed for this problem.Because of the advantages in characteristic learning and feature expression of deep learning in the areas of computer vision and artificial intelligence,we use the deep neural network model to learn the specific statistical features of images to improve the accuracy of forensics.We design a multi-scale module to automatically extract multiple features from the discrete cosine transform(DCT)coefficient histograms of JPEG images.The module can capture characteristic information in different scale spaces,expand and enrich the feature content,and help to improve the effectiveness of subsequent tampering detection.In addition,when the first compression quality factor( 1)of the JPEG image is greater than the second one( 2),due to the minimal statistical difference between the tampered region and the untampered region,the detection effect of the image forensics methods are generally poor.We design a discriminative module for such difficult situations.The specially designed deep neural network in the discriminative module can extract the slight differences between the image DCT coefficient histograms of the tampered region and the real region in such cases,so as to achieve the purpose of improving the tamperdetection effect of the network.Finally,we can obtain a tamper detection probability map and automatically locate the tampered region of the JPEG image.Through a large number of experiments,we have proved that our proposed method is superior to other methods in the quantitative evaluation and visual effect.In order to make a more comprehensive study for the application of multi-scale deep network in double JPEG compression forensics,we conduct further research on another multi-scale deep network structure and propose a multi-scale deep fusion network based algorithm for JPEG image forensics.The algorithm uses a single network to solve the problem of forensics,and the characteristic of multi-scale fusion is mainly reflected in the specific structure of the network.The network can automatically extract multiple features in multi-scale spaces from the DCT coefficient histograms of JPEG image and integrate them effectively to obtain more diversified feature information and expand the content of the extracted features,so as to improve the accuracy of image forensics.Qualitative and quantitative experiments show that the method has a significant increase in subjective and objective evaluation indicators.