Learning to localize image forgery using boundary-preserving mask R-CNN.

Area of Science:

• Computer Vision
• Multimedia Security
• Digital Forensics

Background:

• Digital image manipulation poses significant challenges to multimedia security.
• Existing forgery detection methods often lack precision in localizing manipulated regions, particularly near boundaries, and struggle with generalization across diverse manipulation types.

Purpose of the Study:

• To develop an advanced digital image forgery detection framework.
• To enhance the accuracy and robustness of manipulated region localization, especially near image boundaries.

Main Methods:

• Proposed a novel Boundary-Preserving Mask R-CNN framework.
• Integrated channel attention mechanisms for detailed spatial information capture.
• Utilized frequency domain features for improved robustness.
• Evaluated on six benchmark datasets (CASIA V2, Columbia, Carvalho, CoMoFoD, MICC-F220, CG-1050) covering splicing, copy-move, and compositing manipulations.
• Employed extensive preprocessing and pixel-level segmentation for accurate region detection.

Main Results:

• The Boundary-Preserving Mask R-CNN demonstrated strong performance across multiple evaluation metrics.
• Achieved high accuracy, precision, recall, F1-score, IoU, and AUC.
• Showcased superior localization accuracy, particularly for manipulations near image boundaries.
• Exhibited robustness across various manipulation types and datasets.

Conclusions:

• The proposed framework offers a reliable and precise solution for digital image forgery detection.
• It significantly advances the state-of-the-art in localizing manipulated image regions.
• The method holds substantial potential for applications in digital forensics and security investigations.