Scholarlink

<p>With the rapid advancement of digital image editing technologies, forged images have become increasingly realistic, raising concerns for news credibility, public trust, and judicial forensics. Despite recent progress, existing image tampering localization methods still face challenges in cross-scale semantic alignment, boundary-aware representation, and global structural modeling. To mitigate these issues, this paper presents an Enhanced Gated Attention Network (EGA-Net) for pixel-level image tampering localization. The proposed framework incorporates three coordinated components. First, a Dual-Gated Aggregation module is introduced to align and interact features from consistency-aware and semantic branches through explicit discrepancy modeling. Second, an Edge-Aware Enhancement Block is designed to reinforce multi-scale boundary-related cues and stabilize edge responses under complex visual conditions. Third, a Global Attention Refinement Head integrates semantic representations with structural inconsistency information to refine tampering masks via global context modeling. Extensive experiments on four public benchmark datasets show that EGA-Net achieves consistent improvements over existing methods in terms of F1-score and area under the receiver operating characteristic curve. Notably, the proposed approach exhibits favorable performance in challenging splicing scenarios and cases involving subtle tampering boundaries. These results suggest that EGA-Net provides a reliable and effective framework for image tampering localization. The source code is publicly available at <a href="https://github.com/krtgvz/EGA-Net/tree/master">https://github.com/krtgvz/EGA-Net/tree/master</a>.</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

EGA-Net: An enhanced gated attention network for pixel-level image tampering localization

  • 期刊论文

  • 发表日期：

    2026年3月19日

查看全文
  • Yunxue Shao,
  • Chuanyu Yang

摘要

With the rapid advancement of digital image editing technologies, forged images have become increasingly realistic, raising concerns for news credibility, public trust, and judicial forensics. Despite recent progress, existing image tampering localization methods still face challenges in cross-scale semantic alignment, boundary-aware representation, and global structural modeling. To mitigate these issues, this paper presents an Enhanced Gated Attention Network (EGA-Net) for pixel-level image tampering localization. The proposed framework incorporates three coordinated components. First, a Dual-Gated Aggregation module is introduced to align and interact features from consistency-aware and semantic branches through explicit discrepancy modeling. Second, an Edge-Aware Enhancement Block is designed to reinforce multi-scale boundary-related cues and stabilize edge responses under complex visual conditions. Third, a Global Attention Refinement Head integrates semantic representations with structural inconsistency information to refine tampering masks via global context modeling. Extensive experiments on four public benchmark datasets show that EGA-Net achieves consistent improvements over existing methods in terms of F1-score and area under the receiver operating characteristic curve. Notably, the proposed approach exhibits favorable performance in challenging splicing scenarios and cases involving subtle tampering boundaries. These results suggest that EGA-Net provides a reliable and effective framework for image tampering localization. The source code is publicly available at https://github.com/krtgvz/EGA-Net/tree/master.