Dual-Decoupling With Frequency-Spatial Domains for Image Manipulation Localization.

Area of Science:

• Computer Vision
• Digital Forensics
• Machine Learning

Background:

• Image manipulation localization (IML) relies on trace-rich features in embedded spaces.
• Existing IML methods struggle with redundant information in manipulated trace features.
• This complexity hinders the full comprehension of trace features for accurate localization.

Purpose of the Study:

• To introduce a novel decoupling representation learning network (DRN) for improved image manipulation localization.
• To effectively decouple multidomain information into representations pertinent to IML objectives.
• To enhance the accuracy and robustness of image manipulation detection.

Main Methods:

• Developed a decoupling representation learning network (DRN) for IML.
• Introduced a frequency decoupling module (FDM) to separate low and high-frequency components, reducing redundancy.
• Implemented a spatial decoupling module (SDM) using channel activation maps to distinguish authentic and manipulated representations.

Main Results:

• The proposed DRN method demonstrated superior performance across three public benchmarks (CASIA, NIST, Coverage).
• The network achieved enhanced robustness compared to existing state-of-the-art IML methods.
• Decoupled high-frequency components served as effective trace complements, improving feature aggregation.

Conclusions:

• The DRN effectively addresses the challenge of redundant information in IML by decoupling intricate multidomain features.
• The proposed FDM and SDM modules significantly enhance the precision and reliability of image manipulation localization.
• DRN offers a robust and high-performing solution for identifying manipulated images.