mjkwon2021/CAT-Net

What it solves

CAT-Net is designed to detect and localize manipulated regions in images by analyzing compression artifacts. It addresses the problem of image forgery, specifically targeting splicing (where an object is moved from one image to another) and general image manipulations in its second version.

How it works

The network analyzes JPEG compression artifacts to trace and identify where an image has been altered. It offers two versions:

• CAT-Net v1: Specifically targets splicing forgery.
• CAT-Net v2: Extends the capability to detect general forgery types using mathematical analysis of compression artifacts.

Who it’s for

This project is primarily for image forensics researchers and developers working on multimedia forensics and image manipulation detection.

Highlights

• Two-stage detection: Supports both splicing-specific detection (v1) and general forgery detection (v2).
• Comprehensive dataset support: Provides access to five custom datasets (including tampCOCO and compRAISE) and links to other standard datasets like CASIA and IMD2020.
• Forensic heatmaps: Generates prediction heatmaps to visualize exactly where the manipulated regions are located.
• Pretrained weights: Includes weights for both the v1 and v2 models for immediate inference.