CMV2U-Net: A U-shaped network with edge-weighted features for detecting and localizing image splicing

Area of Science:

• Computer Vision
• Digital Image Forensics
• Deep Learning

Background:

• Image splicing is a common image manipulation technique.
• Current deep learning methods for splicing detection suffer from poor feature fusion and overfitting due to simple models.
• Accurate localization of manipulated image regions is crucial for digital forensics.

Purpose of the Study:

• To propose CMV2U-Net, an improved deep learning approach for image splicing forgery localization.
• To address limitations in feature fusion and model complexity in existing methods.
• To enhance the robustness and accuracy of image manipulation detection.

Main Methods:

• Developed CMV2U-Net, an edge-weighted U-shaped network for image splicing detection.
• Implemented a dual-stream feature extraction module for simultaneous semantic and agnostic feature processing.
• Utilized a hierarchical fusion approach with channel attention to preserve shallow features and monitor manipulation trajectories.

Main Results:

• CMV2U-Net achieved high AUC and F1 scores in localizing tampered regions across multiple public datasets.
• The proposed method demonstrated superior performance compared to state-of-the-art image splicing detection techniques.
• CMV2U-Net proved robust against post-processing threats like noise, Gaussian blur, and JPEG compression.

Conclusions:

• CMV2U-Net offers a significant advancement in image splicing forgery localization.
• The model's architecture effectively overcomes feature fusion and overfitting issues.
• CMV2U-Net provides a reliable and robust solution for detecting manipulated images in real-world scenarios.