融合CBAM的双向递归门控残差图像Copy-Move篡改检测
CSTR:
作者:
作者单位:

(新疆大学 电气工程学院,乌鲁木齐 830017)

作者简介:

马颖(1994—),女,硕士研究生;伊力哈木·亚尔买买提(1978—),男,教授,硕士生导师

通讯作者:

伊力哈木·亚尔买买提,65891080@qq.com

中图分类号:

TP391

基金项目:

国家自然科学基金(62362063);新疆维吾尔自治区自然科学基金(2023D01C21)


Fusion of CBAM with bidirectional recurrent gated residual network for image Copy-Move forgery detection
Author:
Affiliation:

(School of Electrical Engineering, Xinjiang University, Urumqi 830017, China)

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    为解决卷积神经网络的图像Copy-Move篡改检测网络中对小尺度篡改的检测不敏感及细节敏感性低的问题,提出融合卷积块注意力模块(convolutional block attention module,CBAM)的双向递归门控残差图像Copy-Move篡改检测网络。通过两种残差块构建编码器解码器结构,使网络对小尺度篡改区域识别更加敏感;在此基础上,引入CBAM注意力机制,以增强特征表达能力,自适应地聚焦于关键的篡改区域;融合双向门控循环单元(bidirectional gated recurrent unit,Bi-GRU),交互前后向信息,耦合宏微观特征,建模长距离依赖;最后,通过跳跃连接机制,促进多尺度特征融合,强化网络的识别能力。在公共数据集CASIA2.0和CoMoFoD上进行实验,结果表明:本文所提网络在CASIA2.0数据集中的准确率 和F1分数分别达到95.78%和82.08%,验证了其在小尺度篡改检测方面的有效性;在CoMoFoD数据集中的准确率和F1分数分别达到92.36%和55.72%,进一步证实了其对复杂篡改场景的泛化能力。所提网络能准确实现对Copy-Move篡改图像的检测,可为篡改检测领域的后续研究提供技术支撑。

    Abstract:

    To address the issue of insensitivity to small-scale tampering detection and low detail sensitivity in convolutional neural networks for image Copy-Move forgery detection, a bidirectional recurrent gated residual image Copy-Move forgery detection network incorporating the convolutional block attention module (CBAM) is proposed. An encoder-decoder structure is constructed using two types of residual blocks to enhance the network′s sensitivity to identify small-scale tampered regions. Building on this, the CBAM attention mechanism is introduced to strengthen feature representation capabilities and adaptively focus on critical tampered regions. Bidirectional gated recurrent units (Bi-GRU) are incorporated to interact forward and backward information, couple macro and micro features, and model long-range dependencies. Finally, a skip connection mechanism is employed to facilitate multi-scale feature fusion and strengthen the network′s recognition capability. Experiments are conducted on the public datasets CASIA2.0 and CoMoFoD. On the CASIA2.0 dataset, the proposed network achieves an accuracy of 95.78% and an F1-score of 82.08%, validating its effectiveness in small-scale tampered region detection. On the CoMoFoD dataset, it achieves an accuracy of 92.36% and an F1-score of 55.72%, further confirming its generalization capability for complex tampering scenarios. The proposed network can accurately detect Copy-Move forged images, providing technical support for subsequent research in the field of tampering detection.

    参考文献
    相似文献
    引证文献
引用本文

马颖,伊力哈木·亚尔买买提,侯雪扬,程铄棋.融合CBAM的双向递归门控残差图像Copy-Move篡改检测[J].哈尔滨工业大学学报,2026,58(9):154. DOI:10.11918/202412014

复制
分享
相关视频

文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2024-12-04
  • 最后修改日期:
  • 录用日期:
  • 在线发布日期: 2026-10-11
  • 出版日期:
文章二维码