Research Article
Semantic Modeling and Pixel Discrimination for Image Manipulation Detection
Figure 2
The overall architecture of our method. Forgery object-level branch: the features are extracted from the whole image using several convolutional layers. The forgery object information is extracted using encoder to form the feature maps to RPN. Pixel-level branch: the image is divided into 8 × 8 image patches, and the resampling features are extracted for each patch, combined with LSTM to build a temporal relationship. Feature fusion module: the forgery object-level features in the upper branch and the pixel-level features in the lower branch are integrated to produce the heat maps using a decoder. Integrated fusion module: according to the forgery object information of the RPN network, the noises in the heat maps are masked.