Preprint
Computer Vision

Image Manipulation Detection Based on Irrelevant Information Suppression and Critical Information Enhancement

Yunxue Shao(College of Computer and Information Engineering (College of Artificial Intelligence), Nanjing Tech University, Nanjing, Jiangsu Province, China), Tingting Wang(College of Computer and Information Engineering (College of Artificial Intelligence), Nanjing Tech University, Nanjing, Jiangsu Province, China), Lingfeng Wang(Binzhou Institute of Technology, Weiqiao-UCAS Science and Technology Park, Binzhou, Shandong Province, China)
January 29, 2025The European Journal on Artificial Intelligence6 citations
DOI

6

Citations

1

Influential Citations

The European Journal on Artificial Intelligence

Venue

2025

Year

Abstract

The image manipulation detection localization task differs from traditional computer vision tasks in that we focus more on capturing subtle and generic manipulation detection features in images. In this paper, we propose a novel method called irrelevant visual information suppression, which aims to alleviate the interference of irrelevant visual information in images on manipulation detection feature extraction, thereby obtaining generic manipulation traces that are more subtle and unrelated to semantic visual information. In general, most manipulation operations leave traces at manipulation edges. Therefore, we introduce a specially designed manipulated edge information enhancement branch aimed at identifying these edge artifacts more accurately. We construct a dual-branch network, where each branch uses ResNet-50 as the backbone to capture as many multi-scale manipulation features as possible. Finally, we adopt a multi-view feature learning method that combines the manipulated edge information enhancement branch with the irrelevant visual information suppression branch and is trained with multi-scale (pixel/edge/image/irrelevant visual information suppression) supervision. To validate the effectiveness of the proposed method, we conducted extensive experiments using five image manipulation localization datasets, including CASIAv1, CASIAv2, COVER, Columbia, and NIST16. The experimental results demonstrate that our proposed method can outperform state-of-the-art methods by a significant margin in terms of F1 score. Taking CASIAv1, COVER, and Columbia datasets as examples, compared with MVSS-Net published in ICCV 2021, our method has improved F1 scores by 7.1%, 6.3%, and 12.5%, respectively. The code used in this paper can be found at the following URL: https://github.com/ginwins/ISIE-Net .

Analysis

Why This Paper Matters

Image manipulation detection is critical for combating misinformation and digital fraud. Unlike typical computer vision tasks that rely on semantic content, manipulation detection must capture subtle, generic artifacts left by editing operations. This paper addresses a key challenge: irrelevant visual information (e.g., object textures, backgrounds) often distracts models from learning true manipulation traces. By explicitly suppressing such information and enhancing manipulated edge features, the proposed method achieves a notable leap in detection accuracy.

The dual-branch design is particularly significant because it separates two complementary objectives: ignoring irrelevant content and focusing on manipulation-specific edges. This mirrors how forensic experts analyze images—looking past the scene to find inconsistencies at boundaries. The multi-scale supervision further ensures the model learns robust features at pixel, edge, image, and suppression levels, making it more generalizable across different manipulation types and datasets.

Technical Contributions

  • Irrelevant Visual Information Suppression (IVIS): A novel branch that learns to downweight semantic features (e.g., object shapes, colors) that are irrelevant to manipulation detection, forcing the network to focus on subtle, non-semantic traces.
  • Manipulated Edge Information Enhancement (MEIE): A dedicated branch that amplifies edge artifacts—such as blur, double edges, or unnatural transitions—that commonly occur at manipulation boundaries.
  • Dual-Branch Architecture with ResNet-50 Backbones: Each branch uses a strong feature extractor to capture multi-scale manipulation features, then fuses them via multi-view learning.
  • Multi-Scale Supervision: Training with four loss terms (pixel-level, edge-level, image-level, and IVIS-level) ensures comprehensive learning from fine-grained to global manipulation cues.

Results

The method was evaluated on five standard benchmarks: CASIAv1, CASIAv2, COVER, Columbia, and NIST16. Compared to the strong baseline MVSS-Net (ICCV 2021), the proposed ISIE-Net achieved F1 score improvements of 7.1% on CASIAv1, 6.3% on COVER, and 12.5% on Columbia. These are substantial gains in a field where even 1-2% improvement is considered significant. The consistent outperformance across diverse datasets suggests the method generalizes well to different manipulation types (e.g., splicing, copy-move, removal).

Significance

This work pushes the boundary of image forensics by demonstrating that explicitly suppressing irrelevant visual information can dramatically improve manipulation detection. The dual-branch framework is a principled way to decouple content and manipulation features, which could inspire similar approaches in other forensic tasks (e.g., deepfake detection, steganalysis). The public code release (https://github.com/ginwins/ISIE-Net) enables reproducibility and further research. For AI practitioners, this method offers a ready-to-use architecture that can be integrated into automated content verification pipelines, helping combat digital misinformation at scale.

image-manipulation-detectiondeep-learningresnet-50edge-enhancementmulti-scale-supervisiondual-branch-networkcomputer-vision
Computer Vision

Related Papers