| Copy-move forgery is a prevalent form of image forgery in which a specific area of an image is copied and pasted onto another area of the same image.Current deep learning-based methods for detecting copy-paste forgery mainly rely on learning non-descriptive semantic features in the image to distinguish between the source and target regions.However,convolution and downsampling operations can cause the loss of detailed information required for forgery detection.Furthermore,these methods often require feature similarity calculations to locate similar regions in the image,which can result in a significant amount of computation during the detection process,overlooking structural similarities between features.(1)To address the first issue,a dual-branch framework based on multi-scale edge artifacts is proposed.The dual-branch network simultaneously extracts edge artifact-feature and similarity feature,which helps to effectively distinguish the statistical differences between the source and target regions.The top-down fusion module combines multi-scale features to effectively remove redundant noise while preserving edge detail information.The effectiveness of the proposed method was confirmed through ablation experiments.Comparative experiments with current methods further highlighted the superior performance of the proposed method in region localization across multiple datasets.(2)To addresses the problem that existing methods often struggle to effectively distinguish between source and target regions.In the fourth chapter,a novel end-to-end concatenation method is proposed that utilizes Graph Convolutional Network(GCN)to identify differences in feature similarity between the source and target regions.By doing so,the extracted features not only encapsulate the semantic difference between the source and target regions,but also the structural similarity between them.Ablation experiments show the effectiveness of the proposed algorithm,while comparative and robustness experiments confirm its advantages,feasibility,generalization,and robustness in localizing copy-paste manipulations across multiple datasets. |
