F²Mamba: Fine-grained forgery-aware Mamba for image manipulation localization

Area of Science:

• Computer Vision
• Digital Forensics
• Artificial Intelligence

Background:

• Image Manipulation Localization (IML) is vital for detecting digital forgeries and preventing misinformation spread.
• Existing Transformer-based methods for IML excel at global feature modeling but are computationally intensive and neglect crucial local details.
• The need for efficient and accurate pixel-level forgery identification remains a significant challenge in digital forensics.

Purpose of the Study:

• To propose a novel and efficient method, Fine-grained Forgery-aware Mamba (F²Mamba), for accurate image manipulation localization.
• To address the limitations of existing methods by integrating global and local feature analysis effectively.
• To improve the precision and generalization capability of image manipulation localization models.

Main Methods:

• F²Mamba leverages VMamba for efficient, low-cost multi-scale global feature learning.
• A Fine-grained Forgery-aware Adapter (FFA) adaptively fuses local details with global representations for enhanced decoding.
• A Forgery-Guided Refinement Decoder (FRD) with iterative Conditional Random Field (CRF) refinement minimizes boundary errors and artifacts.

Main Results:

• F²Mamba achieved an average F1 score of 69.92% and an average IoU score of 61.26% on diverse datasets.
• The proposed method demonstrates superior performance compared to state-of-the-art IML models.
• Experimental results validate the model's strong generalization capability and effectiveness in precise forgery localization.

Conclusions:

• F²Mamba offers an efficient and accurate solution for image manipulation localization.
• The integration of global modeling and fine-grained local detail fusion significantly improves forgery detection.
• The method shows promise for real-world applications in combating digital misinformation and ensuring public safety.