Digital image manipulation detection with weak feature stream.

With the rapid development of deep neural network, two-stream Faster R-CNN network has been applied to tampering detection field and achieved good detection results. However, the noise stream generation of the two-stream Faster R-CNN needs to be manually selected, and the feature extraction network is the same as RGB stream, so it doesn't maximize the program to play the advantages of deep neural network. In this paper, the hand-crafted features of the Faster R-CNN noise stream generation layer are cancelled and the two-layer convolution network is used to directly fit the weak features generated by image tampering. At the same time, according to the characteristics of image tampering detection, a weak feature extraction network based on multi-scale residual network is established to extract weak feature signals of image tampering. The network can suppress the natural features of the image and preserve the weak feature signals. In the network, the multi-level residual layers are used to extract RoI features, which makes full use of the feature layer informations with higher resolution. The experimental results show that the performance of the Faster R-CNN network with weak feature stream has been improved significantly in F1 score and localization of the tampering region. [ABSTRACT FROM AUTHOR]