Research On Digital Image Content Forgery Localization Algorithms Based On Deep Learning

The rapid development of digital image processing technology has enabled more and more users to modify or synthesize images at a relatively low cost,making them look very "real".Although the intentions for such operations vary widely,malicious image tampering has become a global concern and has had many negative impacts on our lives,such as false news reports,Internet rumors,insurance fraud blackmail,court perjury and even falsification of academic publications and so on.Based on this,image forensics technology is dedicated to identifying tampered images and preventing tampered images from being used for malicious purposes,and has received widespread attention from academia and business circles.Although significant progress has been made in this field,the research on image tampering location is still in its infancy,and most methods focus on one or two types of tampering.At the same time,the handcraft feature in the traditional method combines the features learned by the deep neural network,ignoring the multi-scale information or global information,while only using the traditional handwriting features and ignoring the multi-scale or global information.With the rise and development of deep learning,the field of image forensics will also usher in new development opportunities.In order to further improve the accuracy of the algorithm,this paper focuses on the common image splicing,copy-paste and removal of tampering methods in digital image content change and tampering,and researches on the location of image content tampering based on deep learning.The specific content is as follows:1.Image manipulation detection and localization based on the Dual-Domain Convolutional Neural NetworksBefore the development of deep learning-based methods,traditional methods based on image blocks had a strong dependence on environment,illumination and other features,and ignored more abundant discriminant features such as space and frequency domain.At the same time,in the actual application of image forensics,compared with the task of detecting whether the image is real,the task of locating the tampered area in the tampered image is more important and faces more challenges.In order to solve the above problems,this paper proposes a dual domain-based Convolutional Neural Networks(D-CNNs)based on different types of inputs.In the proposed framework,image blocks in the spatial domain and frequency domain features are used as inputs to the network sub-model separately to extract rich dualdomain discriminative features.Although CNNs have the ability to learn classification features directly from data,their standard form tends to learn features related to image content and ignore tampering trace features in image forensics tasks.To overcome this problem,this paper constructs a new image preprocessing layer that adaptively learns tamper trace features while suppressing image content features.After studying the properties of the dataset,two post-processing operations are proposed and compared to improve the accuracy of tampered region boundary localization.Experiments show that the post-processing operation based on the graph segmentation algorithm can significantly optimize the final tampering probability map,and that the F1 scores of D-CNNs are significantly better than other comparison algorithms on the same dataset.2.PL-GNet: Pixel Level Global Network for Detection and Localization of Image Forgeries based on Atrous ConvolutionMost of the existing image tampering localization algorithms use image blocks as input,which leads to the inability to accurately locate the edge of the image tampering area,which reduces the accuracy of tampering positioning.At the same time,in order to achieve pixel-level positioning results,additional post-processing operations need to be introduced.In order to solve the above problems,the features from the whole images and frequency domain are leveraged in this paper to classify each pixel in the image,and proposes a pixel-level global network called PL-GNet(Pixel Level Global Network)without additional pre-processing and post-processing operations.The end-to-end PL-GNet framework contains three building blocks: 1)An Encoding net based on atrous convolution,which uses atrous convolution to extract rich spatial information and is constructed with multiple parallel atrous convolutional layers with four different rates are constructed to localize tampered regions at multiple scales.2)A Long Short Term Memory(LSTM)network based on co-occurrence is designed to capture the tampering traces and the discriminative features between manipulated and non-manipulated regions.3)A Decoding net which incorporates the output of Encoding Net and LSTM network learns the mapping from low-resolution feature maps to pixel-wise prediction masks.Furthermore,a series of ablation experiments are carefully conducted to systematically optimize the Encoding Net design,and its superior performance was verified.Experimental results prove that,compared with some of the latest mainstream algorithms,PL-GNet performs best on the six standard image datasets.3.Global Semantic Consistency Network for Image Manipulation Detection and LocalizationExisting methods usually ignore semantic differences between different levels of feature maps and directly fuse(e.g.,addition,or concatenation)for feature fusion.In this paper,we believe that the semantic gap is the main reason for the low efficiency of feature fusion in image forgery detection and localization.In order to solve the above problems,we propose a Global Semantic Consistency Network(GSCNet)based on the encoder-decoder structure.Although the classic encoder-decoder structure and its improvements can improve performance to a certain extent,they usually ignore the semantic differences between different levels of feature maps.In order to make full use of the high-resolution information of low-level features in CNN and the abstract semantic information of high-level features,this paper combines the pre-defined resampling features and LSTM model in the encoder-decoder structure to locate the image tampering area.Specifically,to make GSCNet include more global texture information which has been empirically confirmed to be beneficial to manipulation detection,gram block is first deployed on each level of feature maps in the encoding stage.Based on that,Bi-directional Convolutional LSTM(BConv LSTM)is further implemented on the decoding stage,such that feature maps of the same level have semantic consistency.Experimental results on three different datasets show that GSCNet can locate the manipulated regions with high accuracy.Furthermore,compared to the existing models,GSCNet can achieve new state-of-the-art results.