Image Tamper Detection Based on Two-Stream Attention Faster R-CNN

In the past, image tamper detection was mainly based on the traditional way of manually designing features for feature extraction. In many recent studies, image tamper detection based on deep learning has achieved excellent results. We propose an end-to-end image tamper detection model called Two-Stream Attention Net (TSA-Net) based on RGB-N and Faster R-CNN, where the Region of Interest (RoI) pooling in RGB-N is replaced with RoI align. In the training process, by using the constrained convolution layer and steganalysis rich model (SRM) filter to extract the image noise and adding the attention module to the model, our model can significantly enhance the locating of the boundary of RGB image and the noise area where the noise image is inconsistent. In the testing process, in order to reduce false negative, we replaced standard Non-Maximum Suppression (NMS) by using Soft-NMS. Through the use of the attention module and Soft-NMS, the detection ability of the model in small targets is improved. After experiments on four standard datasets, the model significantly surpasses the original model in the detection of image splicing, image copy-move, and image removal.