Research On Image Copy-Move Forgery Detection Based On Convolutional Neural Network

With the development of digital technology,image forgery has become a common phenomenon.A large number of tampered images are spreading rapidly on the Internet,causing great harm to society.Image forgery detection aims to safeguard the authenticity of information,prevent fraud.It will have a positive effect on people’s lives and society.The research object of this paper is image copy-move forgery detection.In the field of image copy-move forgery detection,it is very challenging to accurately locate the boundaries of tampered small objects.At the same time,only individual methods can distinguish the source and target regions.To address the problem of inaccurate boundary recognition for small objects,a detection method based on multi-scale feature extraction and fusion was proposed.Firstly,abundant features were extracted by multi-scale feature extraction and fusion,which is beneficial for detecting tampered objects of different scales.Secondly,skip connection was added between the feature extraction module and the decoding module to bridge the gap between encoded and decoded features,so as to accurately identify the boundaries of small objects.Finally,Log-Cosh Dice Loss was used to take place of cross entropy loss to improve the impact of class-imbalance issue on detection results.In order to distinguish the source region and the target region from the tampered image,a method based on the Siamese network was proposed.Based on the results of the previous similarity detection method,two image blocks were cropped from the original image and used as the input to the Siamese network.The tampering traces of the image blocks were detected by the Siamese network to identify the source and target regions.Experiments on three public datasets(USCISI,CASIA v2.0,CoMoFoD)demonstrate that the proposed method has more accurate boundary identification and better visualization for small objects.At the same time,the source-target region in the detection results can be effectively distinguished.