Research On Image Splicing Forgery Detection Algorithm Based On Deep Learning

With the development of information technology,images,as an important carrier of information transmission,have a wider range of applications.However,the continuous development of image editing software makes it easier to tamper with image content,which brings great harm to social information security.Therefore,the society urgently needs reliable and effective digital image forensics technology.In this paper,related research is carried out on image splicing forgery detection and tampering area location.The main research contents are as follows:Firstly,this paper proposes an image splicing forgery detection algorithm based on noise inconsistency.This algorithm is mainly aimed at the problem that the image splicing forgery detection algorithms based on image segmentation network are easily interfered by the semantic content of the image.It uses the noise inconsistency in the spliced image and uses the noise feature to suppress the influence of the image content on the tampering detection results,and to provide additional evidence.Firstly,the classical image segmentation network UNet is improved,and a path is added to utilize the visual features and noise features existing in the spliced images at the same time,so the model is a two-branch structure.The RGB visual branch is used to extract visual features in the spliced images,such as contrast differences,etc.The Noise branch uses the noise image generated by the SRM filter as input to extract the noise features in the spliced image.Finally,the model fuses the two branches through an improved feature fusion module and outputs the predicted localization.The experimental results show that,compared with the splicing forgery detection scheme based on image segmentation network,the algorithm can achieve more accurate location of tampered regions,and has good robustness to JPEG compression,Gaussian noise,and size transformation.Then,this paper proposes an image splicing forgery detection algorithm based on HRNet and SRM.The algorithm is based on the parallel connection network HRNet,so high-resolution features can be retained all the time,thus solving the defect of the serial connection structure model losing important spatial information such as edge features in the downsampling layers.Firstly,HRNet is improved by replacing ordinary convolution with atrous convolution,thereby reducing the multiple of model downsampling and further preserving rich tampering information.Combined with the SRM technology introduced above,a two-branch network structure is constructed,and at the same time,forensic clues are found from the visual level and the noise level.Through multiple multi-scale feature fusion,the expressive ability of the features is further enhanced,and finally the visual features and noise features of the corresponding resolution are concatenated,and the predicted positioning is output.The experimental results in the CASIA v2.0 dataset and the Columbia dataset show that the algorithm is superior to the existing algorithms,and the localization effect of the tampered area is improved to a certain extent.Finally,this paper proposes an image splicing forgery detection algorithm based on the Attention module.Based on the previous research,this algorithm further explores how to coordinately process visual features and noise features.Considering that the previous algorithm only uses simple feature fusion methods such as concatenate or addition,it is not enough to coordinate multi-class features.Therefore,this paper combines the attention mechanism to design a multi-feature fusion module based on the attribute difference between visual features and noise features.The visual features are processed by the channel attention module,which enhances the attention of the feature patterns that are beneficial for tamper detection,and suppresses the unfavorable features such as image semantic content.The spatial attention module is used to convert noise features into a spatial attention map,which guides the algorithm to look at areas that are more likely to be tampered with.Because this module has the plug-and-play feature,this paper combines it with the image splicing forgery detection algorithm based on noise inconsistency,and proposes an image splicing forgery detection algorithm based on the Attention module.The experimental results show that,compared with the existing algorithm,the algorithm can greatly improve the positioning accuracy of the tampered area.