Scholarlink

<p>The widespread use of malicious image forging necessitates robust methods for precisely localizing image splicing, especially when compression obscures visual indicators of modification. A significant challenge is distinguishing between genuine and manipulated regions utilizing intrinsic signal features such as JPEG compression artifacts. This paper presents an end-to-end fully convolutional network that leverages spatial-domain RGB and discrete cosine transform (DCT) frequency-domain characteristics, implemented with a dual-flow architecture. The RGB flow detects visual artifacts, whereas the DCT flow is pre-trained on double JPEG detection and models compression traces using a binary coefficient representation. To build appropriate localization masks, multiple resolution characteristics are combined. Our experimental validation demonstrates that this approach significantly outperforms state-of-the-art methods, including ManTra-Net. For instance, on the challenging Carvalho dataset, the offered network achieved a mean Intersection over Union (mIoU) of 80.66% on tampered regions, surpassing the 56.29% achieved by the nearest competitor, while simultaneously maintaining an exceptional pixel accuracy of 99.81% on authentic images. Additionally, the DCT sub-flow alone set a new precedent for double JPEG detection, achieving 94.62% accuracy. The current study is the first successful example of combining a multi-resolution approach between the RGB spatial domain and the DCT frequency domain for pixel-level (splicing) localization of image manipulations. This study introduces and establishes a systematic method that provides a general analysis applicable to JPEG- and non-JPEG-formatted media.</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

Hybrid HRNET: An Enhanced Image Manipulation Detection and Localization Based on JPEG Compression

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月10日

查看全文
  • Ahmed Hesham,
  • Ahmed Sedik,
  • Mohammed Elmogy

摘要

The widespread use of malicious image forging necessitates robust methods for precisely localizing image splicing, especially when compression obscures visual indicators of modification. A significant challenge is distinguishing between genuine and manipulated regions utilizing intrinsic signal features such as JPEG compression artifacts. This paper presents an end-to-end fully convolutional network that leverages spatial-domain RGB and discrete cosine transform (DCT) frequency-domain characteristics, implemented with a dual-flow architecture. The RGB flow detects visual artifacts, whereas the DCT flow is pre-trained on double JPEG detection and models compression traces using a binary coefficient representation. To build appropriate localization masks, multiple resolution characteristics are combined. Our experimental validation demonstrates that this approach significantly outperforms state-of-the-art methods, including ManTra-Net. For instance, on the challenging Carvalho dataset, the offered network achieved a mean Intersection over Union (mIoU) of 80.66% on tampered regions, surpassing the 56.29% achieved by the nearest competitor, while simultaneously maintaining an exceptional pixel accuracy of 99.81% on authentic images. Additionally, the DCT sub-flow alone set a new precedent for double JPEG detection, achieving 94.62% accuracy. The current study is the first successful example of combining a multi-resolution approach between the RGB spatial domain and the DCT frequency domain for pixel-level (splicing) localization of image manipulations. This study introduces and establishes a systematic method that provides a general analysis applicable to JPEG- and non-JPEG-formatted media.