Benjamin_Loison/Robust_image_source_identification_on_modern_smartphones
1
0
Fork 0
Code Issues 68 Pull Requests Packages Projects Releases Wiki Activity

Update uncommited

Browse Source
This commit is contained in:
Benjamin Loison 2024-05-23 12:17:54 +02:00
parent 40a88d25c7
commit 7db83a10c9
Signed by: Benjamin_Loison
SSH Key Fingerprint: SHA256:BtnEgYTlHdOg1u+RmYcDE0mnfz1rhv5dSbQ2gyxW8B8
1 changed files with 15 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

30
datasets/raise/fft/remove_period_patterns.py
View File

@ -24,43 +24,43 @@ image = multipleColorsImage
fft1 = fftpack.fftshift(fftpack.fft2(image))
originalFft1 = fft1.copy()
def removePeriodicPatterns(fft1Part):
# This example is intended to demonstrate how astropy.convolve and
# scipy.convolve handle missing data, so we start by setting the brightest
# pixels to NaN to simulate a "saturated" data set
# Unclear why the Gaussian kernel usage does not work as expected, even when just consider the *real* part.
# See [Benjamin_Loison/astropy/issues/2](https://codeberg.org/Benjamin_Loison/astropy/issues/2).
fixedImagePart = fft1Part.copy()
height, width = fft1Part.shape
for x in range(width):
fft1Part[height // 2, x] = np.nan
#fft1Part[height // 2, x] = np.nan
fixedImagePart[height // 2, x] = np.mean([fft1Part[height // 2 - 1, x], fft1Part[height // 2 + 1, x]])
middleX = width // 2
RANGE = 1
for verticalLineX in [width // 4, width // 2, round(3 * width / 4)]:
for y in range(height):
for x in range(verticalLineX - RANGE, verticalLineX + RANGE + 1):
fft1Part[y, x] = np.nan
#fft1Part[y, x] = np.nan
fixedImagePart[y, x] = np.mean([fft1Part[y, x - RANGE - 1], fft1Part[y, x + RANGE + 1]])
# We smooth with a Gaussian kernel
kernel = Gaussian2DKernel(x_stddev = X_STDDEV)
#kernel = Gaussian2DKernel(x_stddev = X_STDDEV)
# create a "fixed" image with NaNs replaced by interpolated values
# See [Benjamin_Loison/astropy/issues/1](https://codeberg.org/Benjamin_Loison/astropy/issues/1).
fixedImagePart = interpolate_replace_nans(fft1Part, kernel)
#fixedImagePart = interpolate_replace_nans(fft1Part, kernel)
return fixedImagePart
# TODO: are `.copy()` really necessary?
realFixedImage = removePeriodicPatterns(np.real(fft1).copy())
imaginaryFixedImage = removePeriodicPatterns(np.imag(fft1).copy())
realFixedImage = removePeriodicPatterns(np.real(fft1))
imaginaryFixedImage = removePeriodicPatterns(np.imag(fft1))
fixedImage = realFixedImage + 1j * imaginaryFixedImage
#plt.imsave('second_image.png', np.log10(1 + abs(fixedImage)))
figure, axes = plt.subplots(1, 2, sharex = True, sharey = True)
plt.suptitle('Attenuating FFT significant lines')
axes[0].set_title('Original FFT')
firstImage = np.log10(1 + abs(originalFft1))
firstImage = np.log10(1 + abs(fft1))
secondImage = np.log10(1 + abs(fixedImage))
images = [firstImage, secondImage]
vMin = np.min(images)
@ -77,18 +77,18 @@ def inverseFft(fft):
ifft2 = np.real(fftpack.ifft2(fftpack.ifftshift(fft)))
return ifft2
invOriginalFft1 = inverseFft(originalFft1)
invFft1 = inverseFft(fft1)
invFixedImage = inverseFft(fixedImage)
images = [invOriginalFft1, invFixedImage]
images = [invFft1, invFixedImage]
vMin = np.min(images)
vMax = np.max(images)
plt.suptitle('Rafael 23/04/24 PRNU mean denoiser with periodic patterns attenuated')
axes[0].set_title('Original PRNU')
axes[0].imshow(invOriginalFft1, vmin = vMin, vmax = vMax)
axes[0].imshow(invFft1, vmin = vMin, vmax = vMax)
axes[1].set_title('PRNU with periodic patterns attenuated')
axes[1].imshow(invFixedImage, vmin = vMin, vmax = vMax)
axes[2].set_title('Difference between both left images')
differenceBetweenBothImages = invOriginalFft1 - invFixedImage
differenceBetweenBothImages = invFft1 - invFixedImage
# `np.log10(1 + abs(differenceBetweenBothImages))` does not seem more interesting.
axes[2].imshow(differenceBetweenBothImages)
plt.tight_layout()