Encoder-decoder based convolutional neural networks for image forgery detection

Today, images editing software has greatly evolved, thanks to them that the semantic manipulation of images has become easier. On the other hand, the identification of these modifications becomes a very difficult task because the modified regions are not visually apparent. In this article, a new convolutional neural network method based on an encoder/decoder called Fals-Unet is proposed to locate the manipulated regions. The encoder of our method uses an architecture topologically identical to that of the Resnet50 method; its main goal is the exploitation of spatial maps to analyze the discriminating characteristics between the manipulated and non-manipulated regions. The decoding network learns the mapping from low-resolution feature maps to pixel-wise predictions for localizing the falsified regions. Finally, the predicted binary mask (0: falsify, 1: not falsify) is generated by the final layer (softmax). Experimental results on many public datasets CASIA, NIST'16, COVERAGE, and COMOD show that the proposed CNN-based model outperforms some methods.