AMontiB
upload
fc8df74
"""
Please read the copyright notice located on the readme file (README.md).
"""
import numpy as np
from scipy import special
import utils.src.Filter as Ft
def crosscorr(array1, array2):
"""
Computes 2D cross-correlation of two 2D arrays.
Parameters
----------
array1 : numpy.ndarray
first 2D matrix
array2: numpy.ndarray
second 2D matrix
Returns
------
numpy.ndarray('float64')
2D cross-correlation matrix
"""
array1 = array1.astype(np.double)
array2 = array2.astype(np.double)
array1 = array1 - array1.mean()
array2 = array2 - array2.mean()
############### End of filtering
normalizator = np.sqrt(np.sum(np.power(array1,2))*np.sum(np.power(array2,2)))
tilted_array2 = np.fliplr(array2); del array2
tilted_array2 = np.flipud(tilted_array2)
TA = np.fft.fft2(tilted_array2); del tilted_array2
FA = np.fft.fft2(array1); del array1
AC = np.multiply(FA, TA); del FA, TA
if normalizator==0:
ret = None
else:
ret = np.real(np.fft.ifft2(AC))/normalizator
return ret
'''
def crosscorr(array1, array2):
# function ret = crosscor2(array1, array2)
# Computes 2D crosscorrelation of 2D arrays
# Function returns DOUBLE type 2D array
# No normalization applied
array1 = array1.astype(np.double)
array2 = array2.astype(np.double)
array1 = array1 - array1.mean()
array2 = array2 - array2.mean()
############### End of filtering
tilted_array2 = np.fliplr(array2); del array2
tilted_array2 = np.flipud(tilted_array2)
TA = np.fft.fft2(tilted_array2); del tilted_array2
FA = np.fft.fft2(array1); del array1
FF = np.multiply(FA, TA); del FA, TA
ret = np.real(np.fft.ifft2(FF))
return ret
'''
def imcropmiddle(X, sizeout, preference='SE'):
"""
Crops the middle portion of a given size.
Parameters
----------
x : numpy.ndarray
2D or 3D image matrix
sizeout: list
size of the output image
Returns
------
numpy.ndarray
cropped image
"""
if sizeout.__len__() >2:
sizeout = sizeout[:2]
if np.ndim(X)==2: X = X[...,np.newaxis]
M, N, three = X.shape
sizeout = [min(M,sizeout[0]), min(N,sizeout[1])]
# the cropped region is off center by 1/2 pixel
if preference == 'NW':
M0 = np.floor((M-sizeout[0])/2)
M1 = M0+sizeout[0]
N0 = np.floor((N-sizeout[1])/2)
N1 = N0+sizeout[1]
elif preference == 'SW':
M0 = np.ceil((M-sizeout[0])/2)
M1 = M0+sizeout[0]
N0 = np.floor((N-sizeout[1])/2)
N1 = N0+sizeout[1]
elif preference == 'NE':
M0 = np.floor((M-sizeout[0])/2)
M1 = M0+sizeout[0]
N0 = np.ceil((N-sizeout[1])/2)
N1 = N0+sizeout[1]
elif preference == 'SE':
M0 = np.ceil((M-sizeout[0])/2)
M1 = M0+sizeout[0]
N0 = np.ceil((N-sizeout[1])/2)
N1 = N0+sizeout[1]
X = X[M0:M1+1,N0:N1+1,:]
return X
def IntenScale(inp):
"""
Scales input pixels to be used as a multiplicative model for PRNU detector.
Parameters
----------
x : numpy.ndarray('uint8')
2D or 3D image matrix
Returns
------
numpy.ndarray('float')
Matrix of pixel intensities in to be used in a multiplicative model
for PRNU.
"""
T = 252.
v = 6.
out = np.exp(-1*np.power(inp-T,2)/v)
out[inp < T] = inp[inp < T]/T
return out
def LinearPattern(X):
"""
Output column and row means from all 4 subsignals, subsampling by 2.
Parameters
----------
x : numpy.ndarray('float')
2D noise matrix
Returns
-------
dict
A dictionary with the following items:
row means as LP.r11, LP.r12, LP.r21, LP.r22 (column vectors)
column means as LP.c11, LP.c12, LP.c21, LP.c22 (row vectors)
numpy.ndarray('float')
The difference between input X and ZeroMean(X); i.e. X-output would be
the zero-meaned version of X
"""
M, N = X.shape
me = X.mean()
X = X-me
#LP = {"r11":[],"c11":[],"r12":[],"c12":[],"r21":[],"c21":[],"r22":[],"c22":[],"me":[],"cm":[]}
LP = dict(r11=[], c11=[], r12=[], c12=[], r21=[], c21=[], r22=[], c22=[], me=[], cm=[])
LP['r11'] = np.mean(X[::2,::2],axis=1)
LP['c11'] = np.mean(X[::2,::2],axis=0)
cm11 = np.mean(X[::2,::2])
LP['r12'] = np.mean(X[::2,1::2],axis=1)
LP['c12'] = np.mean(X[::2,1::2],axis=0)
cm12 = np.mean(X[::2,1::2]) # = -cm Assuming mean2(X)==0
LP['r21'] = np.mean(X[1::2,::2],axis=1)
LP['c21'] = np.mean(X[1::2,::2],axis=0)
cm21 = np.mean(X[1::2,::2]) # = -cm Assuming mean2(X)==0
LP['r22'] = np.mean(X[1::2,1::2],axis=1)
LP['c22'] = np.mean(X[1::2,1::2],axis=0)
cm22 = np.mean(X[1::2,1::2]) # = cm Assuming mean2(X)==0
LP['me'] = me
LP['cm'] = [cm11,cm12,cm21,cm22]
del X
D = np.zeros([M,N],dtype=np.double)
[aa,bb] = np.meshgrid(LP["c11"],LP["r11"],indexing='ij')
D[::2,::2] = aa+bb+me-cm11
[aa,bb] = np.meshgrid(LP["c12"],LP["r12"],indexing='ij')
D[::2,1::2] = aa+bb+me-cm12
[aa,bb] = np.meshgrid(LP["c21"],LP["r21"],indexing='ij')
D[1::2,::2] = aa+bb+me-cm21
[aa,bb] = np.meshgrid(LP["c22"],LP["r22"],indexing='ij')
D[1::2,1::2] = aa+bb+me-cm22
return LP, D
def NoiseExtract(Im,qmf,sigma,L):
"""
Extracts noise signal that is locally Gaussian N(0,sigma^2)
Parameters
----------
Im : numpy.ndarray
2D noisy image matrix
qmf : list
Scaling coefficients of an orthogonal wavelet filter
sigma : float
std of noise to be used for identicication
(recomended value between 2 and 3)
L : int
The number of wavelet decomposition levels.
Must match the number of levels of WavePRNU.
(Generally, L = 3 or 4 will give pretty good results because the
majority of the noise is present only in the first two detail levels.)
Returns
-------
numpy.ndarray('float')
extracted noise converted back to spatial domain
Example
-------
Im = np.double(cv.imread('Lena_g.bmp')[...,::-1]) % read gray scale test image
qmf = MakeONFilter('Daubechies',8)
Image_noise = NoiseExtract(Im, qmf, 3., 4)
Reference
---------
[1] M. Goljan, T. Filler, and J. Fridrich. Large Scale Test of Sensor
Fingerprint Camera Identification. In N.D. Memon and E.J. Delp and P.W. Wong and
J. Dittmann, editors, Proc. of SPIE, Electronic Imaging, Media Forensics and
Security XI, volume 7254, pages # 0I–01–0I–12, January 2009.
"""
Im = Im.astype(float)
M, N = Im.shape
m = 2**L
# use padding with mirrored image content
minpad=2 # minimum number of padded rows and columns as well
nr = (np.ceil((M+minpad)/m)*m).astype(int); nc = (np.ceil((N+minpad)/m)*m).astype(int) # dimensions of the padded image (always pad 8 pixels or more)
pr = np.ceil((nr-M)/2).astype(int) # number of padded rows on the top
prd= np.floor((nr-M)/2).astype(int) # number of padded rows at the bottom
pc = np.ceil((nc-N)/2).astype(int) # number of padded columns on the left
pcr= np.floor((nc-N)/2).astype(int) # number of padded columns on the right
Im = np.block([
[ Im[pr-1::-1,pc-1::-1], Im[pr-1::-1,:], Im[pr-1::-1,N-1:N-pcr-1:-1]],
[ Im[:,pc-1::-1], Im, Im[:,N-1:N-pcr-1:-1] ],
[ Im[M-1:M-prd-1:-1,pc-1::-1], Im[M-1:M-prd-1:-1,:], Im[M-1:M-prd-1:-1,N-1:N-pcr-1:-1] ]
])
# Precompute noise variance and initialize the output
NoiseVar = sigma**2
# Wavelet decomposition, without redudance
wave_trans = Ft.mdwt(Im,qmf,L)
# Extract the noise from the wavelet coefficients
for i in range(1,L+1):
# Horizontal noise extraction
wave_trans[0:nr//2, nc//2:nc], _ = \
Ft.WaveNoise(wave_trans[0:nr//2, nc//2:nc], NoiseVar)
# Vertical noise extraction
wave_trans[nr//2:nr, 0:nc//2], _ = \
Ft.WaveNoise(wave_trans[nr//2:nr, 0:nc//2],NoiseVar)
# Diagonal noise extraction
wave_trans[nr//2:nr, nc//2:nc], _ = \
Ft.WaveNoise(wave_trans[nr//2:nr, nc//2:nc], NoiseVar)
nc = nc//2
nr = nr//2
# Last, coarest level noise extraction
wave_trans[0:nr,0:nc] = 0
# Inverse wavelet transform
image_noise = Ft.midwt(wave_trans,qmf,L)
# Crop to the original size
image_noise = image_noise[pr:pr+M,pc:pc+N]
return image_noise
def Qfunction(x):
"""
Calculates probability that Gaussian variable N(0,1) takes value larger
than x
Parameters
----------
x : float
value to evalueate Q-function for
Returns
-------
float
probability that a variable from N(0,1) is larger than x
float
logQ
"""
if x<37.5:
Q = 1/2*special.erfc(x/np.sqrt(2))
logQ = np.log(Q)
else:
Q = (1/(np.sqrt(2*np.pi)*x))*np.exp(-np.power(x,2)/2)
logQ = -np.power(x,2)/2 - np.log(x)-1/2*np.log(2*np.pi)
return Q, logQ
def rgb2gray1(X):
"""
Converts RGB-like real data to gray-like output.
Parameters
----------
X : numpy.ndarray('float')
3D noise matrix from RGB image(s)
Returns
-------
numpy.ndarray('float')
2D noise matrix in grayscale
"""
datatype = X.dtype
if X.shape[2]==1: G=X; return G
M,N,three = X.shape
X = X.reshape([M * N, three])
# Calculate transformation matrix
T = np.linalg.inv(np.array([[1.0, 0.956, 0.621],
[1.0, -0.272, -0.647],
[1.0, -1.106, 1.703]]))
coef = T[0,:]
G = np.reshape(np.matmul(X.astype(datatype), coef), [M, N])
return G
def Saturation(X, gray=False):
"""
Determines saturated pixels as those having a peak value (must be over 250)
and a neighboring pixel of equal value
Parameters
----------
X : numpy.ndarray('float')
2D or 3D matrix of image with pixels in [0, 255]
gray : bool
Only for RGB input. If gray=true, then saturated pixels in output
(denoted as zeros) result from at least 2 saturated color channels
Returns
-------
numpy.ndarray('bool')
binary matrix, 0 - saturated pixels
"""
M = X.shape[0]; N = X.shape[1]
if X.max()<=250:
if not gray:
SaturMap = np.ones(X.shape,dtype=np.bool)
else:
SaturMap = np.ones([M,N],dtype=np.bool)
return SaturMap
SaturMap = np.ones([M,N],dtype=np.int8)
Xh = X - np.roll(X, 1, axis=1)
Xv = X - np.roll(X, 1, axis=0)
Satur = np.logical_and(np.logical_and(Xh, Xv),
np.logical_and(np.roll(Xh, -1, axis=1),np.roll(Xv, -1, axis=0)))
if np.ndim(np.squeeze(X))==3:
maxX = []
for j in range(3):
maxX.append(X[:,:,j].max())
if maxX[j]>250:
SaturMap[:,:,j] = np.logical_not(np.logical_and(X[:,:,j]==maxX[j],
np.logical_not(Satur[:,:,j])))
elif np.ndim(np.squeeze(X))==2:
maxX = X.max()
SaturMap = np.logical_not(np.logical_and(X==maxX, np.logical_not(SaturMap)))
else: raise ValueError('Invalid matrix dimensions')
if gray and np.ndim(np.squeeze(X))==3:
SaturMap = SaturMap[:,:,1]+SaturMap[:,:,3]+SaturMap[:,:,3]
SaturMap[SaturMap>1] = 1
return SaturMap
def SeeProgress(i):
"""
SeeProgress(i) outputs i without performing carriage return
This function is designed to be used in slow for-loops to show how the
calculations progress. If the first call in the loop is not with i=1, it's
convenient to call SeeProgress(1) before the loop.
"""
if i==1 | i==0 : print('\n ')
print('* %(i)d *' % {"i": i}, end="\r")
def WienerInDFT(ImNoise,sigma):
"""
Removes periodical patterns (like the blockness) from input noise in
frequency domain
Parameters
----------
ImNoise : numpy.ndarray('float')
2D noise matrix extracted from one images or a camera reference pattern
sigma : float
Standard deviation of the noise that we want not to exceed even locally
in DFT domain
Returns
-------
numpy.ndarray('float')
filtered image noise (or camera reference pattern) ... estimate of PRNU
"""
M,N = ImNoise.shape
F = np.fft.fft2(ImNoise); del ImNoise
Fmag = np.abs(np.real(F / np.sqrt(M*N))) # normalized magnitude
NoiseVar = np.power(sigma, 2)
Fmag1, _ = Ft.WaveNoise(Fmag, NoiseVar)
fzero = np.where(Fmag==0); Fmag[fzero]=1; Fmag1[fzero]=0; del fzero
F = np.divide(np.multiply(F, Fmag1), Fmag)
# inverse FFT transform
NoiseClean = np.real(np.fft.ifft2(F))
return NoiseClean
def ZeroMean(X, zType='CFA'):
"""
Subtracts mean from all subsignals of the given type
Parameters
----------
X : numpy.ndarray('float')
2-D or 3-D noise matrix
zType : str
Zero-meaning type. One of the following 4 options: {'col', 'row', 'both', 'CFA'}
Returns
-------
numpy.ndarray('float')
noise matrix after applying zero-mean
dict
dictionary including mean vectors in rows, columns, total mean, and
checkerboard mean
Example
-------
Y,_ = ZeroMean(X,'col') ... Y will have all columns with mean 0.
Y,_ = ZeroMean(X,'CFA') ... Y will have all columns, rows, and 4 types of
odd/even pixels zero mean.
"""
M, N, K = X.shape
# initialize the output matrix and vectors
Y = np.zeros(X.shape, dtype=X.dtype)
row = np.zeros([M,K], dtype=X.dtype)
col = np.zeros([K,N], dtype=X.dtype)
cm=0
# subtract mean from each color channel
mu = []
for j in range(K):
mu.append(np.mean(X[:,:,j], axis=(0,1)))
X[:,:,j] -= mu[j]
for j in range(K):
row[:,j] = np.mean(X[:,:,j],axis=1)
col[j,:] = np.mean(X[:,:,j],axis=0)
if zType=='col':
for j in range(K): Y[:,:,j] = X[:,:,j] - np.tile(col[j,:],(M,1))
elif zType=='row':
for j in range(K): Y[:,:,j] = X[:,:,j] - np.tile(row[:,j],(N,1)).transpose()
elif zType=='both':
for j in range(K): Y[:,:,j] = X[:,:,j] - np.tile(col[j,:],(M,1))
for j in range(K): Y[:,:,j] = X[:,:,j] - np.tile(row[:,j],(N,1)).transpose()# equal to Y = ZeroMean(X,'row'); Y = ZeroMean(Y,'col');
elif zType=='CFA':
for j in range(K): Y[:,:,j] = X[:,:,j] - np.tile(col[j,:],(M,1))
for j in range(K): Y[:,:,j] = X[:,:,j] - np.tile(row[:,j],(N,1)).transpose() # equal to Y = ZeroMean(X,'both');
for j in range(K):
cm = np.mean(Y[::2, ::2, j], axis=(1,2))
Y[::2, ::2, j] -= cm
Y[1::2, 1::2, j] -= cm
Y[::2, 1::2, j] += cm
Y[1::2, ::2, j] += cm
else:
raise(ValueError('Unknown type for zero-meaning.'))
# Linear pattern data:
LP = {}# dict(row=[], col=[], mu=[], checkerboard_mean=[])
LP['row'] = row
LP['col'] = col
LP['mu'] = mu
LP['checkerboard_mean'] = cm
return Y, LP
def ZeroMeanTotal(X):
"""
Subtracts mean from all black and all white subsets of columns and rows
in a checkerboard pattern
Parameters
----------
X : numpy.ndarray('float')
2-D or 3-D noise matrix
Returns
-------
numpy.ndarray('float')
noise matrix after applying ZeroMeanTotal
dict
dictionary of four dictionaries for the four subplanes, each includes
mean vectors in rows, columns, total mean, and checkerboard mean.
"""
dimExpanded = False
if np.ndim(X)==2: X = X[...,np.newaxis]; dimExpanded = True
Y = np.zeros(X.shape, dtype=X.dtype)
Z, LP11 = ZeroMean(X[::2, ::2, :],'both')
Y[::2, ::2, :] = Z
Z, LP12 = ZeroMean(X[::2, 1::2, :],'both')
Y[::2, 1::2,:] = Z
Z, LP21 = ZeroMean(X[1::2, ::2, :],'both')
Y[1::2, ::2,:] = Z
Z, LP22 = ZeroMean(X[1::2, 1::2, :],'both')
Y[1::2, 1::2,:] = Z
if dimExpanded: Y = np.squeeze(Y)
LP = {}# dict(d11=[], d12=[], d21=[], d22=[])
LP['d11'] = LP11
LP['d12'] = LP12
LP['d21'] = LP21
LP['d22'] = LP22
return Y, LP