Scholarlink

<p>In recent years, with the rapid development of image editing technology, the trustworthiness of multimedia data is facing severe challenges, and the security risks caused by image tampering are increasing, which promotes the research of image tampering localization. Traditional image tampering localization methods usually rely on identifying specific forgery traces, such as JPEG artifacts, edge inconsistencies, camera noise, and so on, and localize the tampered regions by cross-entropy loss. With the rise of deep learning, recent studies explore the boundary difference between tampered regions and real regions by adding Convolutional Neural Network and attention mechanism. However, there are two main weaknesses in these methods. First, the existing two-stream networks lack complementary fusion between blocks, and the features of the two branches lack interaction. The second is that they tend to focus on only a few specific forgery artifacts. In fact, the tampered images in real life may be generated by a variety of forgery methods, leaving a variety of tampering traces. To solve these problems, this paper proposes a novel image tampering Location network based on dual-stream dominant multi-step complementary fusion feature and multi-scale attention enhancement feature representation. The network combines the RGB feature and noise feature of the image to accurately identify and locate the possible tampering traces in the image. Specifically, our proposed RGB-Noise Complementary Fusion Module uses the weighting mechanism to complementary correct RGB features and Noise features, and fuses features to compensate for the lack of single-stream information while highlighting important features. In addition, Hierarchical Attention Module processes features at four scales at multiple levels, enabling the network to capture not only local details, but also global structural information, helping the model more accurately locate tampering regions. Experimental results show that the proposed framework achieves significant performance improvement on multiple tamper positioning reference datasets, and demonstrates good generalization ability.</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

Dual-stream network with complementary fusion and hierarchical attention for image tampering localization

  • 期刊论文

  • 发表日期：

    2025年1月28日

查看全文
  • Zhanpeng Mao,
  • Tongwei Lu

摘要

In recent years, with the rapid development of image editing technology, the trustworthiness of multimedia data is facing severe challenges, and the security risks caused by image tampering are increasing, which promotes the research of image tampering localization. Traditional image tampering localization methods usually rely on identifying specific forgery traces, such as JPEG artifacts, edge inconsistencies, camera noise, and so on, and localize the tampered regions by cross-entropy loss. With the rise of deep learning, recent studies explore the boundary difference between tampered regions and real regions by adding Convolutional Neural Network and attention mechanism. However, there are two main weaknesses in these methods. First, the existing two-stream networks lack complementary fusion between blocks, and the features of the two branches lack interaction. The second is that they tend to focus on only a few specific forgery artifacts. In fact, the tampered images in real life may be generated by a variety of forgery methods, leaving a variety of tampering traces. To solve these problems, this paper proposes a novel image tampering Location network based on dual-stream dominant multi-step complementary fusion feature and multi-scale attention enhancement feature representation. The network combines the RGB feature and noise feature of the image to accurately identify and locate the possible tampering traces in the image. Specifically, our proposed RGB-Noise Complementary Fusion Module uses the weighting mechanism to complementary correct RGB features and Noise features, and fuses features to compensate for the lack of single-stream information while highlighting important features. In addition, Hierarchical Attention Module processes features at four scales at multiple levels, enabling the network to capture not only local details, but also global structural information, helping the model more accurately locate tampering regions. Experimental results show that the proposed framework achieves significant performance improvement on multiple tamper positioning reference datasets, and demonstrates good generalization ability.