Image Copy-Move Forgery Detection and Localization Method Based on Sequence-to-Sequence Transformer Structure  

作　　者：Gang Hao Peng Liang Ziyuan Li Huimin Zhao Hong Zhang 

机构地区：[1]School of Computer Science,Guangdong Polytechnic Normal University,Guangzhou,510630,China

出　　处：《Computers, Materials & Continua》2025年第3期5221-5238,共18页计算机、材料和连续体(英文)

基　　金：support from the General Program of the National Natural Science Foundation of China(GrantNo.62072123);Key R&D Initiatives in Guangdong Province(Grant No.2021B0101220006);the Guangdong Provincial Department of Education’s Key Field Projects for Ordinary Colleges and Universities(Grant Nos.2020ZDZX3059,2022ZDZX1012,2023ZDZX1008);Key R&D Projects in Jiangxi Province(Grant No.20212BBE53002);Key R&D Projects in Yichun City(Grant No.20211YFG4270).

摘　　要：In recent years,the detection of image copy-move forgery(CMFD)has become a critical challenge in verifying the authenticity of digital images,particularly as image manipulation techniques evolve rapidly.While deep convolutional neural networks(DCNNs)have been widely employed for CMFD tasks,they are often hindered by a notable limitation:the progressive reduction in spatial resolution during the encoding process,which leads to the loss of critical image details.These details are essential for the accurate detection and localization of image copy-move forgery.To overcome the limitations of existing methods,this paper proposes a Transformer-based approach for CMFD and localization as an alternative to conventional DCNN-based techniques.The proposed method employs a Transformer structure as an encoder to process images in a sequence-to-sequence manner,substituting the feature correlation calculations of previous methods with self-attention computations.This allows the model to capture long-range dependencies and contextual nuances within the image,preserving finer details that are typically lost in DCNN-based approaches.Moreover,an appropriate decoder is utilized to ensure precise reconstruction of image features,thereby enhancing both the detection accuracy and localization precision.Experimental results demonstrate that the proposed model achieves superior performance on benchmark datasets,such as USCISI,for image copy-move forgery detection.These results show the potential of Transformer architectures in advancing the field of image forgery detection and offer promising directions for future research.

关 键 词：CMFD self-attention TRANSFORMER deep convolutional neural networks 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]