Scholarlink

<p>The rapid growth of digital content creation and manipulation has made copy-move forgery—a technique where a portion of an image is copied and pasted within the same image—a prevalent and challenging threat to image authenticity. This paper proposes a hybrid deep learning framework that effectively combines handcrafted features such as Zernike moments and Discrete Cosine Transform (DCT) with modern deep vision architectures to detect copy-move forgeries with high precision. The framework utilizes a Swin Transformer to capture local spatial features and a Vision Transformer (ViT) to model global contextual dependencies across the image. To further enhance spatial relationship modeling, a lightweight Transformer encoder is employed in place of traditional recurrent networks, enabling better parallelism and faster convergence. Preprocessing steps, including denoising, normalization, and patch segmentation, are applied to refine the input, while ensemble output averaging improves model stability and reduces overfitting. A notable strength of the proposed framework is its resilience to JPEG compression artifacts and image noise. This is achieved through the fusion of handcrafted features like Zernike moments and low-frequency DCT coefficients, which remain stable under compression, with transformer-based global-local representations that are less sensitive to pixel-level distortions. Evaluated on benchmark datasets like CASIA, CoMoFoD, and MICC-F220, the suggested model attains a detection accuracy of up to 99.82%, significantly outperforming conventional CNN-based approaches. This hybrid framework provides a robust, scalable solution for digital image forensics, enhancing content verification across domains like journalism, e-commerce, and legal investigations.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Hybrid Deep Learning Framework for Robust Copy-Move Forgery Detection in Digital Images

  • 期刊论文

  • 发表日期：

    2026年1月22日

查看全文
  • Ibrahim Alfadli,
  • Bala Dhandayuthapani Veerasamy,
  • Santosh Reddy Addula,
  • K S Nandini Prasad,
  • Prashant Kumar Shukla,
  • Weam M. Binjumah

摘要

The rapid growth of digital content creation and manipulation has made copy-move forgery—a technique where a portion of an image is copied and pasted within the same image—a prevalent and challenging threat to image authenticity. This paper proposes a hybrid deep learning framework that effectively combines handcrafted features such as Zernike moments and Discrete Cosine Transform (DCT) with modern deep vision architectures to detect copy-move forgeries with high precision. The framework utilizes a Swin Transformer to capture local spatial features and a Vision Transformer (ViT) to model global contextual dependencies across the image. To further enhance spatial relationship modeling, a lightweight Transformer encoder is employed in place of traditional recurrent networks, enabling better parallelism and faster convergence. Preprocessing steps, including denoising, normalization, and patch segmentation, are applied to refine the input, while ensemble output averaging improves model stability and reduces overfitting. A notable strength of the proposed framework is its resilience to JPEG compression artifacts and image noise. This is achieved through the fusion of handcrafted features like Zernike moments and low-frequency DCT coefficients, which remain stable under compression, with transformer-based global-local representations that are less sensitive to pixel-level distortions. Evaluated on benchmark datasets like CASIA, CoMoFoD, and MICC-F220, the suggested model attains a detection accuracy of up to 99.82%, significantly outperforming conventional CNN-based approaches. This hybrid framework provides a robust, scalable solution for digital image forensics, enhancing content verification across domains like journalism, e-commerce, and legal investigations.