Scholarlink

<p>The increasing complexity of digital image manipulation has rendered content authentication a significant issue in multimedia forensics. To enhance detection systems, it is essential to go beyond just semantic-level clues and also identify subtle, low-level inconsistencies that can help distinguish between authentic and altered content. Traditional methods that combine manually designed features with CNN-based or hybrid models often lack the necessary flexibility or granularity. Convolutional neural networks focus on fundamental semantics, frequently overlooking sophisticated indicators of forgery, whereas manually crafted approaches require an understanding of manipulation styles, which limits their scalability. This study tackles this limitation by introducing an innovative multi-perspective detection framework that separates the image analysis process into three complementary visual pathways. The initial module examines high-frequency components, where manipulation artefacts are frequently observed. The second focuses on non-semantic cues, such as residuals and noise patterns, whereas the third investigates spatial context to enhance overall comprehension. Every branch utilizes ResNet18 for feature enhancement, supplemented by global-axis connections to achieve thorough representation learning. Furthermore, we implement edge supervision through a Sobel-based edge extraction module to improve the precision of boundary detection and manipulation localization. The proposed method attains an F1 score of up to 98.26%, demonstrating a significant advantage over current leading baselines. This architecture focuses on edge-guided learning and multi-channel feature fusion, offering an efficient and adaptable solution for manipulation localization.</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

Multi-perspective CNN framework with edge-guided supervision for robust image manipulation localization

  • 期刊论文

  • 发表日期：

    2026年5月20日

查看全文
  • Monica Luthra,
  • Mahesh Kumar,
  • Rama Kant,
  • Aarti Sehwag,
  • Sonika Jindal,
  • Thejo Lakshmi Gudipalli,
  • G. Murugesan

摘要

The increasing complexity of digital image manipulation has rendered content authentication a significant issue in multimedia forensics. To enhance detection systems, it is essential to go beyond just semantic-level clues and also identify subtle, low-level inconsistencies that can help distinguish between authentic and altered content. Traditional methods that combine manually designed features with CNN-based or hybrid models often lack the necessary flexibility or granularity. Convolutional neural networks focus on fundamental semantics, frequently overlooking sophisticated indicators of forgery, whereas manually crafted approaches require an understanding of manipulation styles, which limits their scalability. This study tackles this limitation by introducing an innovative multi-perspective detection framework that separates the image analysis process into three complementary visual pathways. The initial module examines high-frequency components, where manipulation artefacts are frequently observed. The second focuses on non-semantic cues, such as residuals and noise patterns, whereas the third investigates spatial context to enhance overall comprehension. Every branch utilizes ResNet18 for feature enhancement, supplemented by global-axis connections to achieve thorough representation learning. Furthermore, we implement edge supervision through a Sobel-based edge extraction module to improve the precision of boundary detection and manipulation localization. The proposed method attains an F1 score of up to 98.26%, demonstrating a significant advantage over current leading baselines. This architecture focuses on edge-guided learning and multi-channel feature fusion, offering an efficient and adaptable solution for manipulation localization.