Scholarlink

<p>Image tampering localization techniques aim to identify manipulated regions in digital images. Existing methods rely on specific tampering cues and suffer from poor adaptability and insufficient generalization capabilities. Drawing upon the concept of contrastive learning, our method eliminates the need for a priori knowledge, instead leveraging solely the statistical disparities between real and tampered regions within the feature space. Concurrently, tampering detection independent of specific cues is achieved by mapping the image to a feature domain amplifying the differences in regional features. We compute the comparison loss for each sample to enhance the model's capability of mining the unique tampering features within the samples and improving generalizability. Experiments demonstrate that TB-IFLCL demonstrates good performance on the CASIA, NIST16, and Columbia datasets. On the CASIA dataset, its F1 score is 0.699 and the AUC is 0.875.Through comparative experiments, significant differences are found in the F1 score and AUC between TB-IFLCL and other similar models. This indicates that the model has certain advantages in tampering detection, while its generalization ability is also improved to some extent.</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

Image tampering detection based on two-branch feature extraction and regional feature discrepancy contrastive learning

  • 期刊论文

  • 发表日期：

    2025年7月4日

查看全文
  • Huanhuan Lei,
  • Weiyi Wei,
  • Jian Shi,
  • FuTong Zhang

摘要

Image tampering localization techniques aim to identify manipulated regions in digital images. Existing methods rely on specific tampering cues and suffer from poor adaptability and insufficient generalization capabilities. Drawing upon the concept of contrastive learning, our method eliminates the need for a priori knowledge, instead leveraging solely the statistical disparities between real and tampered regions within the feature space. Concurrently, tampering detection independent of specific cues is achieved by mapping the image to a feature domain amplifying the differences in regional features. We compute the comparison loss for each sample to enhance the model's capability of mining the unique tampering features within the samples and improving generalizability. Experiments demonstrate that TB-IFLCL demonstrates good performance on the CASIA, NIST16, and Columbia datasets. On the CASIA dataset, its F1 score is 0.699 and the AUC is 0.875.Through comparative experiments, significant differences are found in the F1 score and AUC between TB-IFLCL and other similar models. This indicates that the model has certain advantages in tampering detection, while its generalization ability is also improved to some extent.