Copy-Move Forgery Detection Based on PatchMatch

Authors: Marien Renaud, Gonzague de Carpentier

This repository contains a Python implementation of the algorithm described in “Copy-move forgery detection based on PatchMatch”, Cozzolino et al., 2014 [4]. The image database used to test the algorithm comes from the authors' website. All code of the repository is authored by Marien Renaud and Gonzague de Carpentier.

Usage

The main file of the directory is patchmatch.py. It is a stand-alone file (it can be moved to any directory and will still work) and has as only dependencies numpy, matplotlib and numba. It contains the class PatchMatch that enables to run the PatchMatch algorithm. Here is an example of how to use it.

import numpy as np
import matplotlib.pyplot as plt
from PIL import Image

import patchmatch as pm

# Load example image
im = Image.open("data/CMFD_DB/TP_C02_007_copy.png")
im = np.array(im).astype("double")

# Instantiate PatchMatch and compute Zernike moments
a = pm.PatchMatch(
    im,  # image
    p=10,  # patch half-size
    max_zrd=6,  # maximum Zernike degree
    min_dn=64,   # minimum displacement norm (previously T)
    n_rs_candidates=5,   # number of candidates in the random search phase (previously L)
    init_method=2,  # whether to use create_vect_field1 or create_vect_field2
    zernike=True  # whether to use Zernike moments
)

# Run 5 PatchMatch iterations
a.run(5)

# Get resulting displacement field
print(a.vect_field.shape)

Other examples can be found in the jupyter notebooks of the notebooks directory.

Other Python files

The file postprocessing.py contains functions to post-process the displacement field computed by PatchMatch and translate it into a mask of detected copy-moved areas.

The file evaluate.py is a script to evaluate PatchMatch on the entire test database located at data/CMFD_DB/ and store the results in a results/ directory. This script is designed to be easily parallelizable on several cores and several machines. The results can than be loaded in the notebooks to make some plots out of them.

Notebooks

• individual_experiments.ipynb runs PatchMatch on some individual examples and plots the results.
• zernike_moments.ipynb plots the Zernike filters computed by the PatchMatch class.
• evaluation_on_database.ipynb plots results on test database generated by evaluate.py.
• fscore_evolution.ipynb plots the evolution of the F-measure along the iterations of PatchMatch. It loads the results generated by evaluate.py.

References

[1] Connelly Barnes et al. “PatchMatch: A Randomized Correspondence Algorithm for Structural Image Editing”. 2009.

[2] Connelly Barnes et al. “The Generalized PatchMatch Correspondence Algorithm”. 2010.

[3] Davide Cozzolino et al. “Copy-move forgery detection based on PatchMatch”. Oct. 2014.

[4] Davide Cozzolino et al. “Efficient Dense-Field Copy–Move Forgery Detection”. Nov. 2015.

[5] Thibaud Ehret. “Automatic Detection of Internal Copy-Move Forgeries in Images”. July 25, 2018.

[6] Amir Tahmasbi et al. “Classification of benign and malignant masses based on Zernike moments”. Aug. 2011.

[7] Yongqing Xin et al. “Accurate Computation of Zernike Moments in Polar Coordinates”. Feb. 2007.