Research On Detection Algorithms Based On Digital Image Content Tampering

With the development of information technology,images have become an important vehicle for conveying information in digital media.At the same time,the emergence of image editing techniques has made it easy for tamperers to manipulate the contents of images.Therefore,identifying the authenticity of digital images and achieve tampering locality has become an urgent problem that needs to be solved.Image tampering detection algorithms can be divided into traditional methods and deep learning-based methods.Traditional methods require manually designing features that fit the purpose of forensics based on the tampering traces,which is not only difficult but also most algorithms are only targeted at a single tampering method and lack generality.The development of deep learning has provided new ideas for passive forensics,where neural networks can learn deeper features from tampered datasets through their powerful representation capabilities for tampering detection.However,relying solely on neural networks to learn features often results in the loss of edge information and other tampering traces during continuous convolution operations,meanwhile,most neural networks do not explore more information from multiple scales,which can lead to problems such as blurred edge localization,poor generalization,and difficulty in improving detection accuracy.This thesis focuses on how to improve the detection accuracy of image tampering detection algorithms and achieve more precise tampering localization.The algorithm studied in this thesis mainly targets three types of image tampering techniques: copypaste,splicing,and removal.Firstly,the traditional edge detection operator is introduced on top of deep learning to enhance the neural network’s focus on tampering with edge information,thereby improving the algorithm’s localization accuracy.Then,by utilizing full-scale skip connections and deep supervision structure to fuse features from multiple scales,the algorithm’s detection accuracy,generalization,and robustness were improved.The specific research contents are as follows:(1)To address the problems of blurred edges and lost tampering details in current image tampering detection algorithms,a layered edge enhancement image tampering detection scheme is designed.Using RRU-Net(The Ringed Residual U-Net,RRU-Net)as the baseline model,we first introduce the Sobel edge detection operator into the shallow structure of the model to extract the edge information of each layer’s feature.Then,we fuse the extracted edge information with the feature map of that layer to increase the weight of edge information in the feature map.Secondly,the edge information from each layer is output and fused to form an edge feature map,and the label map is used as a guidance for the network parameter update through a designed joint loss function.The experimental results show that the improvement algorithm has some improvement in detection accuracy and localization capability,but the generalization of the method is poor.(2)Designed a image tampering detection scheme based on full-scale deep supervision to address the issues of poor generalization and low detection accuracy in current image tampering detection algorithms.To fully utilize the edge and texture information in the shallow layers that are beneficial to tampering detection,we first design a full-scale skip connection based on the attention gate mechanism on top of the model proposed in Chapter 3.This structure can fully fuse the features from both deep and shallow layers and uses attention gate to enable the network to focus on the tampered regions.Then,side outputs were designed for each upsampling stage to predict the result image,and a joint loss function was designed to optimize the model using deep supervision.Experimental results show that the proposed method in this chapter not only improves the generalization and robustness of the model,but also accurately locates the tampered area with high accuracy on four standard datasets.