Research On Image Tampering Detection Algorithm Based On Deep Learning

As an important carrier for people to obtain and transmit information,digital images not only bring convenience to people's lives,but also bring some security risks to the society.With the development and popularization of image editing software,the cost of image content modification and image forgery has become lower and lower.Users can easily use image editing tools to modify image content without leaving obvious visual traces.The wanton dissemination of tampering images will have a negative impact on the military,politics,media and other fields.Therefore,the research of digital image forensics has very important practical significance.In this paper,we mainly focus on the image manipulation tampering detection and image tampering region localization.The main works are as follow:For image manipulation tampering detection,after the image processing operation,the content of the image itself has not changed.The model needs to learn the noise characteristics of the image,not the content information of the image.The process of traditional algorithms is more complicated,generally requiring manual extraction of features and researchers must have rich knowledge in the field of digital images.This paper proposes an image manipulation detection algorithm based on SRM and residual network.A pre-processing layer is designed which use SRM filter that are proposed in the field of steganalysis and constrained convolution kernel to extract the noise features of the image.The model also contains a multi-scale convolutional layer and residual modules.The multi-scale convolutional layer can fuse feature map of different receptive field sizes.The residual module effectively reduces the risk of gradient disappearance during the model training by establishing short connections between the input and output feature maps,and the convergence speed of the network is accelerated.The experimental results show that the algorithm proposed in this paper can effectively detect five image manipulation types such as mean filtering,median filtering,Gaussian blur,JPEG compression,and contrast enhancement.Aiming at the problem of image tampering region localization,due to the scale diversity of the tampering area and the interference of the blur operation,it is difficult to extract the tampering features of the image with traditional methods.In order to effectively solve this problem,this paper proposes a multi-task image splicing tampering localization algorithm based on Deep Labv3+,using the automatic feature learning and accurate object recognition capabilities of the full convolutional network,convert the multi-object classification problem in Deep Labv3+ to the location problem of tampering area.The model integrates the spatial and channel attention mechanisms in the atrous spatial pyramid pooling to strengthens the model's sensitivity to important features.Furthermore,a branch network that use shallow image feature to predict the boundary of the tampering area is constructed,which enhance the model's attention to the boundary of the tampering area.The experimental results on CASIA V1.0 and Columbia datasets show that the algorithm proposed in this paper can effectively locate the tampering area and assist the identification of the tampered image,the F1-Score and MCCScore are better than other comparison algorithms,the recall of tampered images detection has also been greatly improved.In order to simplify the model's training process and reduce the complexity of the model,this paper uses a light-weight convolutional network to construct an image splicing tampering localization algorithm based on Mobile Net V2 and SRM.The algorithm adjusts the light-weight convolutional network Mobile Netv2,reducing its down-sampling multiple to retain richer traces of image tampering,and uses atrous convolution to enlarge the receptive field of the feature map.In addition,a dual-stream network is constructed in the model,the RGB stream is used to extract the strong contrast difference of the image,the unnatural tampering boundary and other tampering features.Then,the SRM stream is used to extract the noise features between the tampering area and the authentic area of the image.Finally,the extracted dualstream features are fused through the improved attention mechanism.The experimental results show that the network model designed in this paper has high positioning accuracy while achieving light-weight.The localization accuracy in the CASIA v1.0 dataset also better than other comparison algorithms.