FCR-CNN Model to Improve Selection Performance of Image Manipulation Detection Region

Image manipulation detection is different from object detection as it pays more attention to tampered artifacts rather than image content and needs to learn to obtain more abundant features.The FCR-CNN model that combines the feature pyramid networks(FPN)model and the cascade region-convolutional neural networks(Cascade R-CNN)model is proposed to improve the performance of image manipulation detection.The FCR-CNN model first inputs the multi-scale manipulation features extracted by the FPN network to the region proposal network(RPN),then outputs the manipulation classification score and the region proposal from the RPN network,and finally the region proposal is input to the three-stage Cascade R-CNN model for detection.Additionally,the loss function of the FCR-CNN model is systematically analyzed.Comparative experiments with other detection models are also performed on the CASIA,Columbia,and NC2016 datasets.The results show that the FCR-CNN model can effectively detect and locate manipulation areas.Among them,compared with the FPN model and the Cascade R-CNN model on the CASIA dataset,the F1 score of the FCR-CNN model increased by 6.0%and 7.5%,respectively.