An approach based on object detection for image forensics

With the development of the Internet and advanced image editing software, the manipulation and propagation of digital image has become easy, and image forensics has become a challenging research area. In order to detect digital image forgery, various digital image forensics techniques have been proposed. Compared to active forensics methods that require embedded additional information, passive forensics methods are more popular because of their wider application scenarios. In this paper, we propose an approach on the basis of the Faster R-CNN network which has a good performance in object detection. There are two channels in our approach. One channel for extracting noise features is additionally added to find the noise inconsistency between the real area and the tampering area. The other original Faster R-CNN network also serves as a channel for extracting features from the original image to find evidence of tampering such as strong contrast differences and unnatural tampering boundaries. We then fuse the features from the two channels through the bilinear collection layer to further detect image tampering traces. Our experiments on four standard image tamper datasets show that the two-channel framework is superior to the single detection network and has better performance than alternative methods that have the flexibility of resizing and compression in the past.