Scholarlink

<p>With the rapid advancement of digital image processing technology, image splicing forgery has emerged as a prevalent manipulation technique, posing significant threats to information authenticity. While in existing detection methods, semantic information often interferes with feature extraction, thus affecting detection performance. Therefore, this paper proposes a novel image splicing forgery detection model termed JAMD-Net, which based on JPEG compression artifacts and multi-dilated channel refinement fusion. The model employs a dual-branch structure to separately extract JPEG compression artifacts and RGB visual features respectively and fuse them. Specifically, the JPEG Artifact Learning Block is introduced to mine JPEG compression artifact information, effectively avoiding the interference of image semantic content; the Mamba Linear Attention Block uses a special linear attention mechanism to accurately measure the correlation between pixels for successfully capturing long-distance dependency relationships; the Multi-Dilation Channel Refinement Fusion Block refines and fuses features from multiple scales and dimensions, significantly enhancing the feature expression ability. Experimental results show that this model achieves superior performance.</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

JAMD-Net: image splicing forgery detection based on JPEG compression artifacts and multi-dilated channel refinement fusion

  • 期刊论文

  • 发表日期：

    2026年1月4日

查看全文
  • Weiyi Wei,
  • Jian Shi,
  • Huan Huan Lei,
  • Chenglin Chai

摘要

With the rapid advancement of digital image processing technology, image splicing forgery has emerged as a prevalent manipulation technique, posing significant threats to information authenticity. While in existing detection methods, semantic information often interferes with feature extraction, thus affecting detection performance. Therefore, this paper proposes a novel image splicing forgery detection model termed JAMD-Net, which based on JPEG compression artifacts and multi-dilated channel refinement fusion. The model employs a dual-branch structure to separately extract JPEG compression artifacts and RGB visual features respectively and fuse them. Specifically, the JPEG Artifact Learning Block is introduced to mine JPEG compression artifact information, effectively avoiding the interference of image semantic content; the Mamba Linear Attention Block uses a special linear attention mechanism to accurately measure the correlation between pixels for successfully capturing long-distance dependency relationships; the Multi-Dilation Channel Refinement Fusion Block refines and fuses features from multiple scales and dimensions, significantly enhancing the feature expression ability. Experimental results show that this model achieves superior performance.