Research map / Detection and segmentation / Weakly supervised detection and segmentation
M²RL-Net: Multi-View and Multi-Level Relation Learning Network for Weakly-Supervised Image Forgery Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes M2RL-Net, a weakly-supervised image forgery detection network that uses RGB and noise views with a SegFormer encoder, patch-level self-consistency learning, and feature-level contrastive learning, trained only on image-level labels.
Problem
Detecting and localizing image forgeries using only image-level labels, without pixel-level masks, while generalizing to unseen manipulations.
Approach
Combines RGB and learnable-constrained-conv noise views in a SegFormer encoder, mines patch-level self-consistency from transformer attention maps with head-recalibrated relation maps, and uses prototype-based self-view and multi-view contrastive learning on reliable pseudo-labeled pixels.
Key results
- Average image-level I-F1 of 0.762 and AUC of 0.862 across test sets, versus 0.580 I-F1 for WSCL (MIL-FCN+WSCL) baseline per the text's comparison.
- Average pixel-level P-F1 of 0.265 across five datasets, compared with 0.198 for Araslanov+WSCL and 0.419 for fully-supervised MVSS-Net.
- Combined F1 (Com-F1) average of 0.441, outperforming weakly supervised WSCL re-trained (0.267) and fully supervised MVSS-Net (0.474 is higher, so not best overall).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use M2RL-Net instead of the strongest listed weakly supervised baseline | Weakly supervised image forgery detection · CASIAv1 | AUC: 0.836 -> 0.948 | Helped |
| Use M2RL-Net instead of the strongest listed weakly supervised baseline | Weakly supervised image forgery detection · Columbia | I-F1: 0.802 -> 0.975 | Helped |
| Use M2RL-Net instead of the strongest listed weakly supervised baseline | Weakly supervised image forgery detection · COVER | AUC: 0.729 -> 0.716 | Hurt |
| Use M2RL-Net instead of the strongest listed weakly supervised baseline | Weakly supervised image forgery detection · IMD20 | I-F1: 0.449 -> 0.585 | Helped |
| Use M2RL-Net instead of the strongest listed weakly supervised baseline | Weakly supervised image forgery detection · CASIAv1 | P-F1: 0.172 -> 0.347 | Helped |
| Use M2RL-Net instead of the strongest listed weakly supervised baseline | Weakly supervised image forgery detection · Columbia | P-F1: 0.362 -> 0.434 | Helped |
| Use M2RL-Net instead of the strongest listed weakly supervised baseline | Weakly supervised image forgery detection · COVER | P-F1: 0.201 -> 0.213 | Helped |
| Use M2RL-Net instead of the strongest listed weakly supervised baseline | Weakly supervised image forgery detection · IMD20 | P-F1: 0.193 -> 0.248 | Helped |
| Replace RGB view with noise view (ablation) | M2RL-Net · CASIAv1 | I-F1: 0.720 -> 0.634 | Hurt |
| Add multi-view features (ablation) | M2RL-Net · CASIAv1 | AUC: 0.823 -> 0.916 | Helped |
| Add auxiliary classification loss (ablation) | M2RL-Net with multi-view features · CASIAv1 | I-F1: 0.813 -> 0.856 | Helped |
| Add patch-level self-consistency learning (ablation) | M2RL-Net with auxiliary loss · COVER | P-F1: 0.162 -> 0.186 | Helped |
Surprising findings
- Fully supervised MVSS-Net still achieves higher average pixel-level P-F1 (0.419) and Com-F1 (0.474) than the proposed weakly supervised method's 0.265 and 0.441 in the reported tables.
- Several fully supervised baselines (ManTra-Net, SPAN, H-LSTM) collapse to near-chance image-level results under the cross-dataset protocol.
Limitations
- Pseudo-localization labels from approximate maps may be inaccurate, making prototypes and contrastive targets noisy.
- Pixel-level P-F1 remains below fully supervised MVSS-Net on several datasets.
Experimental setup
Datasets: CASIAv2
Benchmarks: CASIAv1 Columbia COVER IMD20 NIST16
Metrics: AUC specificity sensitivity I-F1 P-F1 C-F1
Baselines: NOI1 CFA1 H-LSTM ManTra-Net RRU-Net CR-CNN GSR-Net SPAN CAT-Net FCN+DA MVSS-Net MIL-FCN
Builds on: WSCL SegFormer MVSS-Net EdgeCAM
Closest papers by meaning
- Reinforced Multi-teacher Knowledge Distillation for Efficient General Image Forgery Detection and Localization (AAAI 2025 · similarity 0.87)
- SUMI-IFL: An Information-Theoretic Framework for Image Forgery Localization with Sufficiency and Minimality Constraints (AAAI 2025 · similarity 0.87)
- Identity-Aware Vision-Language Model for Explainable Face Forgery Detection (AAAI 2026 · similarity 0.86)
- Learning Discriminative Noise Guidance for Image Forgery Detection and Localization (AAAI 2024 · similarity 0.86)
- Forensics-Bench: A Comprehensive Forgery Detection Benchmark Suite for Large Vision Language Models (CVPR 2025 · similarity 0.85)
- FakeShield: Explainable Image Forgery Detection and Localization via Multi-modal Large Language Models (ICLR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.