← all work classical-computer-vision

Copy-move forgery

[![Python 3.10+](https://img.shields.io/badge/python-3.10+-3776AB?logo=python&logoColor=white)](https://www.python.org/)

Result

Sr	Scene	Block matching	SIFT + similarity	ORB + similarity	SIFT + translation	SIFT blobs (no verify)
1	penguin on pebbles · every stone resembles every stone	1.000	0.855	0.854	0.851	0.748
2	fighter jet · man-made straight edges	1.000	0.873	0.868	0.860	0.789
3	elephant herd · genuinely repeated objects	1.000	0.861	0.845	0.848	0.836
4	tortoise on rock · broken rock texture	1.000	0.843	0.661	0.842	0.798

The finding

What it does

`mermaid flowchart LR A[Clean photo] --> B["Copy a TEXTURED patch
rotate / rescale it
paste it elsewhere"] B --> C[Exact mask of BOTH copies] B --> D1[Block matching] B --> D2[SIFT / ORB self-match] D2 --> E{Verify what
hypothesis?} E -->|nothing| F1[Paint blobs] E -->|a shift| F2[Translation verify] E -->|shift + rotate + scale| F3[Similarity verify] D1 & F1 & F2 & F3 --> G[Score vs BOTH copies] C --> G A2[UNTAMPERED photos] --> D1 & F1 & F2 & F3 D1 & F1 & F2 & F3 --> H[False alarms:
what does it accuse
a clean photo of?]

style C fill:#fef3c7,stroke:#d97706 style E fill:#dbeafe,stroke:#2563eb style H fill:#fee2e2,stroke:#dc2626 `

Two design choices carry the whole project:

The truth mask marks both copies (amber). After the paste, the source and the destination are *identical* — same content, same noise, same compression. Nothing in the image says which one came first. A detector that finds a duplicated pair has done its job; deciding which half is the forgery needs information the pixels do not contain (lighting direction, perspective, a second photograph). Scoring against the pasted region alone caps precision at about 0.5 for every method and measures that ambiguity instead of the detector.

Untampered images are scored too (red). A detector that marks part of every photograph you feed it is not a detector. On a clean image IoU, precision and recall are all undefined — the truth is empty — so the only honest number is the fraction of pixels flagged, and it should be zero. This is where the rotation-robust methods stop looking free.

---

Figures

Elsewhere

Code and full results on GitHub (code private)

Every number is read from the file the run produced. github.com/hammasbuilds · kaggle.com/muhammadhammas13