Design And Implementation Of Deep Learning Network For Image Tampering Detection And Localization

Image tampering has brought a great negative impact on society,people who don't know the truth can easily be misled and used by those who are interested.Its impact on society has attracted the attention of scholars at home and abroad.With the development of image editing technology,from the initial visual recognition to machine learning to identify tampered images,to the excellent characteristics of deep learning algorithms in processing images in complex image tampering detection in recent years,image tampering detection has achieved impressive results.Eye-catching results.Image tampering technology goes through three stages: 1.Initially,it is realized by using discrete cosine transform or color interpolation matrix technology to calculate image blocks.Although part of the tampered image can be detected,the effect on the compressed or smoothed image is not obvious..2.Tamper detection technology based on machine learning,such as decision tree,vector machine,clustering algorithm,realizes the detection and location of image tampering area.3.A series of improved models based on convolutional neural networks to detect image tampering areas,such as VGG16,the combination of convolutional neural networks and long-and short-term memory networks,etc.,using a large number of data sets to repeatedly train to achieve accurate positioning of the tampering area.The existing tamper detection methods mostly use the idea of relying on the inconsistency of the edge of the detection area,which makes the quality of image tampering detection completely dependent on the inconsistency of the area edge.If the edge of the tampered area is processed,it is likely that the tamper will not be detected.area.Therefore,this topic uses a dual-branch U-Net deep learning framework,which has the following three innovations: 1.According to the different noise characteristics of the tampered area and the non-tampered area,the network branch of noise extraction is proposed.2.In order to enhance the memory capacity of the model,a residual network is added.In order to prevent the model from learning and forgetting during the learning process,a residual feedback network is also proposed.3.In order to enhance the learning ability of the network structure,the attention mechanism is adopted as the optimization goal.The attention mechanism will make the network tend to learn,thereby inhibiting the learning of non-key areas,greatly improving the efficiency and accuracy of visual information processing.The data set used in the model of this subject is CASIAv2.0.In order to ensure the practicality of the model,the data set not only includes people,animals and plants,but also various natural scenery and buildings of various countries.In addition,two data sets,Columbia and Nist,are used,both of which are close to nature and life.Among them,Columbia is a data set taken and synthesized with a camera.The Nist dataset is a high-resolution dataset.Through comparative experiments with U-Net,RRU-Net and other image tampering detection technologies on the same data set,the results show that the tampering detection effect of this model has been significantly improved.