Scholarlink

<p>The expanding quantity of digital editing tools has made picture manipulation a major threat to multimedia content validity. This work introduces a lightweight, robust picture forgery detection methodology for edge and mobile devices. A dual-pathway architecture integrating visual cues in RGB space and edge information in LAB space fused with a multi-scale fusion module improves detection efficiency across manipulation scales. Comprehensive evaluations on CASIA v1, Columbia, IMD2020, and NIST2016 show the model’s superiority in F1-scores, IoU, and AUC values over state-of-the-art techniques. The proposed model achieves an F1-score of 92.50% on CASIA v1, 90.10% on Columbia, and 93.30% on IMD2020, outperforming several state-of-the-art approaches while maintaining a lightweight architecture with only 35 MB memory usage and 0.3 s inference time per image. Real-time applications can use the lightweight design’s low calculation overhead, memory utilisation, and processing speed. Ablation experiments also demonstrate the importance of dual routes and multiscale fusion for optimal performance. These first findings imply that the suggested model is an interesting solution to accurate and fast image forgery detection with bright future prospects in digital and industrial applications.</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

A lightweight dual-pathway model with multiscale fusion for robust image forgery detection

  • 期刊论文

  • 发表日期：

    2026年5月2日

查看全文
  • Shivnarayan Ahirwar,
  • Alpana Pandey,
  • Ramji Gupta

摘要

The expanding quantity of digital editing tools has made picture manipulation a major threat to multimedia content validity. This work introduces a lightweight, robust picture forgery detection methodology for edge and mobile devices. A dual-pathway architecture integrating visual cues in RGB space and edge information in LAB space fused with a multi-scale fusion module improves detection efficiency across manipulation scales. Comprehensive evaluations on CASIA v1, Columbia, IMD2020, and NIST2016 show the model’s superiority in F1-scores, IoU, and AUC values over state-of-the-art techniques. The proposed model achieves an F1-score of 92.50% on CASIA v1, 90.10% on Columbia, and 93.30% on IMD2020, outperforming several state-of-the-art approaches while maintaining a lightweight architecture with only 35 MB memory usage and 0.3 s inference time per image. Real-time applications can use the lightweight design’s low calculation overhead, memory utilisation, and processing speed. Ablation experiments also demonstrate the importance of dual routes and multiscale fusion for optimal performance. These first findings imply that the suggested model is an interesting solution to accurate and fast image forgery detection with bright future prospects in digital and industrial applications.