Scholarlink

<p>In recent years, image forgery technology has attracted increasing attention, especially in social media and digital media. The detection of forged images is crucial for maintaining information security and social trust. Although existing methods have made some progress in detecting forged images, they still face two key problems. First, there is a feature fusion gap between traditional Discrete Cosine Transform (DCT) and deep learning frameworks, which makes it difficult to fully explore the potential features of images. Second, the existing spatial attention mechanisms struggle to capture the periodic artifact characteristics of generative models in the frequency domain. To this end, this paper proposes a forged image detection method based on a two-stream frequency-space joint constraint network, aiming to achieve image forgery detection across generative models. This method first maps features to the frequency domain through a DCT for convolution filtering and then inversely transforms them back to the spatial domain. Subsequently, the two-stream features are fused to generate attention weights, which optimize the feature representation and thereby achieve a balance between capturing global frequency patterns and preserving local details. Secondly, the Conditional Random Field module is used as a post-processing stage to perform structured optimization through an inter-pixel energy function. This further models contextual relationships and enhances the ability of the generative model to capture image structure and semantic consistency. Finally, a comprehensive evaluation index is used to assess the performance of the model. The experimental results on the dataset, which includes 15 generative models (7 GANs and 8 diffusion models) such as ProGAN, StyleGAN, CycleGAN, Stargan, BigGAN, DALL-E, GLIDE, etc., show that the proposed method has relative advantages in detection accuracy and average precision. These results verify its effectiveness and robustness and provide a better solution for the image forgery detection task.</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 FDS-CRF net: frequency-spatial constrained attention for cross-generative model image detection

  • 期刊论文

  • 发表日期：

    2025年7月30日

查看全文
  • Tao Yang,
  • Qian Zhang

摘要

In recent years, image forgery technology has attracted increasing attention, especially in social media and digital media. The detection of forged images is crucial for maintaining information security and social trust. Although existing methods have made some progress in detecting forged images, they still face two key problems. First, there is a feature fusion gap between traditional Discrete Cosine Transform (DCT) and deep learning frameworks, which makes it difficult to fully explore the potential features of images. Second, the existing spatial attention mechanisms struggle to capture the periodic artifact characteristics of generative models in the frequency domain. To this end, this paper proposes a forged image detection method based on a two-stream frequency-space joint constraint network, aiming to achieve image forgery detection across generative models. This method first maps features to the frequency domain through a DCT for convolution filtering and then inversely transforms them back to the spatial domain. Subsequently, the two-stream features are fused to generate attention weights, which optimize the feature representation and thereby achieve a balance between capturing global frequency patterns and preserving local details. Secondly, the Conditional Random Field module is used as a post-processing stage to perform structured optimization through an inter-pixel energy function. This further models contextual relationships and enhances the ability of the generative model to capture image structure and semantic consistency. Finally, a comprehensive evaluation index is used to assess the performance of the model. The experimental results on the dataset, which includes 15 generative models (7 GANs and 8 diffusion models) such as ProGAN, StyleGAN, CycleGAN, Stargan, BigGAN, DALL-E, GLIDE, etc., show that the proposed method has relative advantages in detection accuracy and average precision. These results verify its effectiveness and robustness and provide a better solution for the image forgery detection task.