Copy-move Detection Method Based On Conditional Generative Adversarial Networks

With the rapid development of digital devices,images have gradually become an important carrier for disseminating visual information.However,with the advent of image editing software,ordinary users even without professional knowledge of image processing can easily modify images without leaving obvious visual traces.Copy-move forgery is a common method of image tampering.It copies an area in an image,and after zooming or rotating the area,move it to other positions of the same image to achieve the purpose of hiding important information or forging false scenes.In recent years,in order to prevent copy-move from being used in illegal crimes,the technique of copy-move forgery detection has developed rapidly,playing an active role in maintaining social order and information security.In order to to solve the problem of low accuracy of existing methods for copy-move forgery detection,we propose an image copy-move forgery localization method based on Conditional Generating Adversation Networks(CGAN).The proposed method can effectively identify whether the digital image has been forged by copy-move or not.Furthermore,the proposed method can localize the similar regions and even distinguish the source and the target regions in copy-move images.The proposed CGANs based copy-move localization method improves the accuracy of localization performance and provides more reliable image forensics information.The main research contents of this paper are as follows:1.Aiming at copy-move forged image and non-forged image data,a classification network based on xception is proposed to realize the effective distinction between copy-move forged image samples and non-forged image samples.2.For copy-move forged images and their corresponding ground truth data pairs,based on conditional generative adversarial networks(CGAN),a novel method is pro-posed for the localization of copy-move forgery,which can not only localize the similar regions in copy-move forged images,but also further distinguish the forged source region and target region.According to the characteristics of copy-move forgery detection tasks,this thesis proposes two optimization methods.On the one hand,the loss function of the network model is optimized.In addition to the adversarial loss,the traditional loss based on L1 distance and the L.ask loss which measures the error between the forgery source and target region are introduced.On the other hand,this thesis proposes a training method that uses a certain number of weakly supervised samples,namely non-forged image samples,to improve the performance of the model.Non-forged image samples refer to the samples in which all areas in the image have not been tampered with,but they can provide a single type of supervision information.Existing methods usually ignore such samples,but they also carry supervision information.Appropriate and reasonable use can effectively improve model performance.Extensive experimental results show that the proposed method remarkably outperforms the compared methods in localization accuracy.