Estimating fake PRNU on noise-free images #25

New Issue

Open

opened 2024-03-28 15:59:48 +01:00 by Benjamin_Loison · 16 comments

Benjamin_Loison commented 2024-03-28 15:59:48 +01:00

Owner

Copy Link

Related to #19, #27 and #9.

 Related to #19, #27 and #9.

prnu.png

11 KiB

Benjamin_Loison added the

high priority

epic

labels 2024-03-28 15:59:48 +01:00

Benjamin_Loison referenced this issue 2024-03-28 15:59:57 +01:00

Estimating PRNU on Gaussian noise images #21

Benjamin_Loison pinned this 2024-03-28 16:00:41 +01:00

Benjamin_Loison commented 2024-03-29 00:34:14 +01:00

Author

Owner

Copy Link

As Context-Adaptive Interpolator is slow and possibly incorrectly implemented, let us try a good known working denoiser:

import matplotlib.pyplot as plt

from skimage.restoration import (denoise_tv_chambolle, denoise_bilateral,
                                 denoise_wavelet, estimate_sigma)
from skimage import data, img_as_float
from skimage.util import random_noise


original = img_as_float(data.chelsea()[100:250, 50:300])

sigma = 0.155
noisy = random_noise(original, var=sigma**2)

fig, ax = plt.subplots(nrows=2, ncols=4, figsize=(8, 5),
                       sharex=True, sharey=True)

plt.gray()

# Estimate the average noise standard deviation across color channels.
sigma_est = estimate_sigma(noisy, channel_axis=-1, average_sigmas=True)
# Due to clipping in random_noise, the estimate will be a bit smaller than the
# specified sigma.
print(f'Estimated Gaussian noise standard deviation = {sigma_est}')

tiles = [
    {
        'title': 'Original',
        'image': original,
    },
    {
        'title': 'Noisy',
        'image': noisy,
    },
    {
        'title': 'TV',
        'image': denoise_tv_chambolle(noisy, weight=0.1, channel_axis=-1),
    },
    {
        'title': '(more) TV',
        'image': denoise_tv_chambolle(noisy, weight=0.2, channel_axis=-1),
    },
    {
        'title': 'Bilateral',
        'image': denoise_bilateral(noisy, sigma_color=0.05, sigma_spatial=15, channel_axis=-1),
    },
    {
        'title': '(more) Bilateral',
        'image': denoise_bilateral(noisy, sigma_color=0.1, sigma_spatial=15, channel_axis=-1),
    },
    {
        'title': 'Wavelet denoising',
        'image': denoise_wavelet(noisy, channel_axis=-1, rescale_sigma=True),
    },
    {
        'title': 'Wavelet denoising\nin YCbCr colorspace',
        'image': denoise_wavelet(noisy, channel_axis=-1, convert2ycbcr=True, rescale_sigma=True),
    },
]

for y in range(2):
    for x in range(4):
        axis = ax[y, x]
        tile = tiles[y * 4 + x]
        axis.imshow(tile['image'])
        axis.axis('off')
        rms = rmsDiffNumpy(tile['image'], original)
        axis.set_title(tile['title'] + f'\nRMS with original: {round(rms, 4)}')

fig.tight_layout()

plt.show()

Based on https://scikit-image.org/docs/0.22.x/auto_examples/filters/plot_denoise.html

As Context-Adaptive Interpolator is slow and possibly incorrectly implemented, let us try a good known working denoiser:  ```py import matplotlib.pyplot as plt from skimage.restoration import (denoise_tv_chambolle, denoise_bilateral, denoise_wavelet, estimate_sigma) from skimage import data, img_as_float from skimage.util import random_noise original = img_as_float(data.chelsea()[100:250, 50:300]) sigma = 0.155 noisy = random_noise(original, var=sigma**2) fig, ax = plt.subplots(nrows=2, ncols=4, figsize=(8, 5), sharex=True, sharey=True) plt.gray() # Estimate the average noise standard deviation across color channels. sigma_est = estimate_sigma(noisy, channel_axis=-1, average_sigmas=True) # Due to clipping in random_noise, the estimate will be a bit smaller than the # specified sigma. print(f'Estimated Gaussian noise standard deviation = {sigma_est}') tiles = [ { 'title': 'Original', 'image': original, }, { 'title': 'Noisy', 'image': noisy, }, { 'title': 'TV', 'image': denoise_tv_chambolle(noisy, weight=0.1, channel_axis=-1), }, { 'title': '(more) TV', 'image': denoise_tv_chambolle(noisy, weight=0.2, channel_axis=-1), }, { 'title': 'Bilateral', 'image': denoise_bilateral(noisy, sigma_color=0.05, sigma_spatial=15, channel_axis=-1), }, { 'title': '(more) Bilateral', 'image': denoise_bilateral(noisy, sigma_color=0.1, sigma_spatial=15, channel_axis=-1), }, { 'title': 'Wavelet denoising', 'image': denoise_wavelet(noisy, channel_axis=-1, rescale_sigma=True), }, { 'title': 'Wavelet denoising\nin YCbCr colorspace', 'image': denoise_wavelet(noisy, channel_axis=-1, convert2ycbcr=True, rescale_sigma=True), }, ] for y in range(2): for x in range(4): axis = ax[y, x] tile = tiles[y * 4 + x] axis.imshow(tile['image']) axis.axis('off') rms = rmsDiffNumpy(tile['image'], original) axis.set_title(tile['title'] + f'\nRMS with original: {round(rms, 4)}') fig.tight_layout() plt.show() ``` Based on https://scikit-image.org/docs/0.22.x/auto_examples/filters/plot_denoise.html

Figure_1.svg

826 KiB

Benjamin_Loison commented 2024-03-29 00:43:02 +01:00

Author

Owner

Copy Link

With PRNU_FACTOR = 0.1:

PRNU quite guessable on a given image. Now need to reduce its factor and mean on several images to make it clear.

With PRNU_FACTOR = 0.01 it is more subtile but still visible on both images. For the image with PRNU see the U. For the PRNU estimate look at the top of R and bottom of U for instance.

PRNU estimate taking into account all images:

Manipulating brightness and contrast on GIMP does not enable a clear result as the following one. However, the not significant RMS decreasing reason is unclear. Even with normalizing it does not seem to make sense.

With `PRNU_FACTOR = 0.1`: `PRNU` quite guessable on a given image. Now need to reduce its factor and mean on several images to make it clear.   With `PRNU_FACTOR = 0.01` it is more subtile but still visible on both images. For the image with PRNU see the `U`. For the PRNU estimate look at the top of `R` and bottom of `U` for instance.    PRNU estimate taking into account all images:  Manipulating brightness and contrast on GIMP does not enable a clear result as the following one. However, the not significant RMS decreasing reason is unclear. Even with normalizing it does not seem to make sense.

Figure_1.svg

1003 KiB

Figure_1.svg

269 KiB

Figure_1.svg

1002 KiB

Figure_1.svg

279 KiB

Figure_1.svg

250 KiB

Figure_1.svg

376 KiB

Figure_1.svg

991 KiB

Benjamin_Loison commented 2024-03-29 01:15:17 +01:00

Author

Owner

Copy Link

Split images in 4x4 to increase PRNU estimation accuracy:

 Split images in 4x4 to increase PRNU estimation accuracy: 

Figure_1.svg

111 KiB

prnu_4x4.png

9.2 KiB

Benjamin_Loison commented 2024-03-29 01:48:12 +01:00

Author

Owner

Copy Link

Let us not consider this RMS inconsistency but admit that the visual clear PRNU is fine, so let us add Gaussian noise.

Let us not consider this RMS inconsistency but admit that the visual clear PRNU is fine, so let us add Gaussian noise.

Benjamin_Loison commented 2024-03-29 01:54:29 +01:00

Author

Owner

Copy Link

In a first time we show an example of Gaussian noise and PRNU on an image, then we consider the PRNU estimate on the mean of the extracted noise for the considered images: in a first time only full images and in a second time split each images in 4x4.

With NOISE_FACTOR = 0.25:

With NOISE_FACTOR = 0.1:

In a first time we show an example of Gaussian noise and PRNU on an image, then we consider the PRNU estimate on the mean of the extracted noise for the considered images: in a first time only full images and in a second time split each images in 4x4. With `NOISE_FACTOR = 0.25`:    With `NOISE_FACTOR = 0.1`:   

Figure_2.svg

1.8 MiB

Figure_1.svg

1.1 MiB

Figure_1.svg

143 KiB

Figure_1.svg

466 KiB

Figure_1.svg

1.0 MiB

Figure_1.svg

142 KiB

Benjamin_Loison commented 2024-03-29 02:03:38 +01:00

Author

Owner

Copy Link

If want to move on Gaussian noise as PRNU, then have to understand RMS unexpected results. May also be surprised, so can try anyway. To have recognizable PRNU can use GIMP > Filters > Noise > RGB Noise... selecting white and setting {Red,Blue,Green} to 1.

If want to move on Gaussian noise as PRNU, then have to understand RMS unexpected results. May also be surprised, so can try anyway. To have recognizable PRNU can use `GIMP` > `Filters` > `Noise` > `RGB Noise...` selecting white and setting `{Red,Blue,Green}` to `1`.    

image.png

284 KiB

prnu_4x4_noise.png

33 KiB

Figure_1.svg

1.8 MiB

Figure_1.svg

128 KiB

Benjamin_Loison commented 2024-03-29 11:22:50 +01:00

Author

Owner

Copy Link

Concerning the RMS, maybe it is due to residual scene and with more images we would have a many pixels, hence important, less significant as nearer to 0.

Concerning the RMS, maybe it is due to residual scene and with more images we would have a many pixels, hence important, less significant as nearer to 0.

Benjamin_Loison commented 2024-03-29 11:26:05 +01:00

Author

Owner

Copy Link

How to ease visualizing the values of this rendering of image of floats? Could render as a greyscale image to make sure of my rough understanding of image values.

How to ease visualizing the values of this rendering of image of floats? Could render as a greyscale image to make sure of my rough understanding of image values.

Benjamin_Loison commented 2024-03-29 12:19:30 +01:00

Author

Owner

Copy Link

 

Figure_1.svg

869 KiB

Figure_1.svg

240 KiB

Benjamin_Loison commented 2024-03-29 13:39:01 +01:00

Author

Owner

Copy Link

I start having the feeling that the RMS inconsistency is because of the PRNU estimation background which can not converge to 0 as it is just adding residual scene.

I start having the feeling that the RMS inconsistency is because of the PRNU estimation background which can not converge to 0 as it is just adding residual scene.

Benjamin_Loison commented 2024-03-29 13:47:04 +01:00

Author

Owner

Copy Link

The associated code is at dd7a6892e5/datasets/noise_free_test_images/estimate_prnu.py.

 The associated code is at https://gitea.lemnoslife.com/Benjamin_Loison/Robust_image_source_identification_on_modern_smartphones/src/commit/dd7a6892e50a6fb4ef9142028d9b25ab835e6598/datasets/noise_free_test_images/estimate_prnu.py.

Benjamin_Loison commented 2024-03-29 14:24:46 +01:00

Author

Owner

Copy Link

Next step:

Try on known working old cameras, i.e. with actual noise and PRNU? See #29.

Next step: Try on known working old cameras, i.e. with actual noise and PRNU? See #29.

Benjamin_Loison closed this issue 2024-03-30 00:08:18 +01:00

Benjamin_Loison added the

enhancement

label 2024-03-30 00:08:52 +01:00

Benjamin_Loison referenced this issue 2024-04-02 10:59:29 +02:00

Comparing noises on images, are there clever distances? #5

Benjamin_Loison referenced this issue 2024-04-08 04:44:20 +02:00

Given device split into two groups and compare both estimated PRNUs and how they evolved when consider more and more images to estimate them within each group #31

Benjamin_Loison unpinned this 2024-04-23 04:50:43 +02:00

Benjamin_Loison referenced this issue 2024-04-30 03:38:26 +02:00

Optimize execution speed #62

Benjamin_Loison commented 2024-07-04 02:40:48 +02:00

Author

Owner

Copy Link

Add a single or multiple color scales?

Related to Benjamin_Loison/matplotlib/issues/32.

Based on the Stack Overflow question 23876588:

Using:

axis[0].set_title('First image without noise')
axis0 = axis[0].imshow(imageWithoutPrnuNpArrayTile)
plt.colorbar(axis0, label = 'Intensity', ax = axis[0])

With tight_layout:

Without tight_layout:

With cax instead of ax:

The Stack Overflow answer 49037495 code leads to:

/home/benjamin/.local/lib/python3.10/site-packages/matplotlib/projections/__init__.py:63: UserWarning: Unable to import Axes3D. This may be due to multiple versions of Matplotlib being installed (e.g. as a system package and as a pip package). As a result, the 3D projection is not available.
  warnings.warn("Unable to import Axes3D. This may be due to multiple versions of "
Traceback (most recent call last):
  File "<tmp 1>", line 14, in <module>
    cax = divider.append_axes('right', size='5%', pad=0.05)
  File "/home/benjamin/.local/lib/python3.10/site-packages/matplotlib/_api/deprecation.py", line 387, in wrapper
    return func(*inner_args, **inner_kwargs)
  File "/usr/lib/python3/dist-packages/mpl_toolkits/axes_grid1/axes_divider.py", line 552, in append_axes
    ax = self.new_horizontal(size, pad, pack_start=False, **kwargs)
  File "/usr/lib/python3/dist-packages/mpl_toolkits/axes_grid1/axes_divider.py", line 486, in new_horizontal
    ax = self._get_new_axes(**kwargs)
  File "/usr/lib/python3/dist-packages/mpl_toolkits/axes_grid1/axes_divider.py", line 437, in _get_new_axes
    axes_class = axes._axes_class
AttributeError: 'Axes' object has no attribute '_axes_class'

The Stack Overflow answer 23877042 code leads to:

Traceback (most recent call last):
  File "<tmp 2>", line 162, in <module>
    plt.register_cmap(cmap=blue_red2)
AttributeError: module 'matplotlib.pyplot' has no attribute 'register_cmap'

The other answers to the mentioned Stack Overflow question do not seem interesting.

orientation = 'horizontal' makes the figure ok.

Add a single or multiple color scales? Related to [Benjamin_Loison/matplotlib/issues/32](https://codeberg.org/Benjamin_Loison/matplotlib/issues/32). Based on [the Stack Overflow question 23876588](https://stackoverflow.com/q/23876588): Using: ```python axis[0].set_title('First image without noise') axis0 = axis[0].imshow(imageWithoutPrnuNpArrayTile) plt.colorbar(axis0, label = 'Intensity', ax = axis[0]) ``` With `tight_layout`:  Without `tight_layout`:  With `cax` instead of `ax`:  [The Stack Overflow answer 49037495](https://stackoverflow.com/a/49037495) code leads to: ``` /home/benjamin/.local/lib/python3.10/site-packages/matplotlib/projections/__init__.py:63: UserWarning: Unable to import Axes3D. This may be due to multiple versions of Matplotlib being installed (e.g. as a system package and as a pip package). As a result, the 3D projection is not available. warnings.warn("Unable to import Axes3D. This may be due to multiple versions of " Traceback (most recent call last): File "<tmp 1>", line 14, in <module> cax = divider.append_axes('right', size='5%', pad=0.05) File "/home/benjamin/.local/lib/python3.10/site-packages/matplotlib/_api/deprecation.py", line 387, in wrapper return func(*inner_args, **inner_kwargs) File "/usr/lib/python3/dist-packages/mpl_toolkits/axes_grid1/axes_divider.py", line 552, in append_axes ax = self.new_horizontal(size, pad, pack_start=False, **kwargs) File "/usr/lib/python3/dist-packages/mpl_toolkits/axes_grid1/axes_divider.py", line 486, in new_horizontal ax = self._get_new_axes(**kwargs) File "/usr/lib/python3/dist-packages/mpl_toolkits/axes_grid1/axes_divider.py", line 437, in _get_new_axes axes_class = axes._axes_class AttributeError: 'Axes' object has no attribute '_axes_class' ``` [The Stack Overflow answer 23877042](https://stackoverflow.com/a/23877042) code leads to: ``` Traceback (most recent call last): File "<tmp 2>", line 162, in <module> plt.register_cmap(cmap=blue_red2) AttributeError: module 'matplotlib.pyplot' has no attribute 'register_cmap' ``` The other answers to the mentioned Stack Overflow question do not seem interesting. `orientation = 'horizontal'` makes the figure ok.

with_tight_layout.svg

192 KiB

without_tight_layout.svg

398 KiB

cax_with_tight_layout.svg

334 KiB

Benjamin_Loison reopened this issue 2024-07-04 02:40:48 +02:00

Benjamin_Loison commented 2024-07-04 04:16:42 +02:00

Author

Owner

Copy Link

Related to Benjamin_Loison/matplotlib/issues/35.

Related to [Benjamin_Loison/matplotlib/issues/35](https://codeberg.org/Benjamin_Loison/matplotlib/issues/35).

Benjamin_Loison commented 2024-07-11 22:53:08 +02:00

Author

Owner

Copy Link

Should check the denoiser range. There is no denoiser here.
Could use the scikit util view as block/window.

Should check the denoiser range. There is no denoiser here. Could use the scikit util view as block/window.

Benjamin_Loison referenced this issue 2024-07-17 14:00:41 +02:00

Track `matplotlib`: Adapt the color scale to leverage it the most when zoom #75

Benjamin_Loison changed title from Estimating fake PRNU on noise-less images to Estimating fake PRNU on noise-free images 2024-07-23 22:43:16 +02:00

Benjamin_Loison commented 2024-07-24 14:18:37 +02:00

Author

Owner

Copy Link

Related to PRNU_extraction/issues/34.

Related to [PRNU_extraction/issues/34](https://codeberg.org/Benjamin_Loison/PRNU_extraction/issues/34).

Sign in to join this conversation.

No Branch/Tag Specified

Branches Tags

master

prnu_written_as_such

No results found.

Labels

Clear labels   

bug

Something is not working

  

Context-Adaptive Interpolator

  

duplicate

This issue or pull request already exists

  

enhancement

New feature

  

epic

A task that is going to take more than a day to complete.

  

help wanted

Need some help

  

high priority

  

invalid

Something is wrong

  

left for future work

  

low priority

  

medium

A task that should take less than a day to complete.

  

medium priority

  

meta

  

question

More information is needed

  

quick

A task that should take less than two hours to complete.

  

wontfix

This won't be fixed

No Label

bug

Context-Adaptive Interpolator

duplicate

enhancement

epic

help wanted

high priority

invalid

left for future work

low priority

medium

medium priority

meta

question

quick

wontfix

Milestone

Clear milestone

No items

No Milestone

Projects

Clear projects

No project

Assignees

Clear assignees

No Assignees

1 Participants

Notifications

Subscribe

Due Date

The due date is invalid or out of range. Please use the format 'yyyy-mm-dd'.

No due date set.

Dependencies

No dependencies set.

Reference: Benjamin_Loison/Robust_image_source_identification_on_modern_smartphones#25

No description provided.