| 1 |
""" |
| 2 |
.. _image_fft_example: |
| 3 |
|
| 4 |
Fast Fourier Transform |
| 5 |
~~~~~~~~~~~~~~~~~~~~~~ |
| 6 |
|
| 7 |
This example shows how to apply a Fast Fourier Transform (FFT) to a |
| 8 |
:class:`pyvista.ImageData` using :func:`pyvista.ImageDataFilters.fft` |
| 9 |
filter. |
| 10 |
|
| 11 |
Here, we demonstrate FFT usage by denoising an image, effectively removing any |
| 12 |
"high frequency" content by performing a `low pass filter |
| 13 |
<https://en.wikipedia.org/wiki/Low-pass_filter>`_. |
| 14 |
|
| 15 |
This example was inspired by `Image denoising by FFT |
| 16 |
<https://scipy-lectures.org/intro/scipy/auto_examples/solutions/plot_fft_image_denoise.html>`_. |
| 17 |
|
| 18 |
""" |
| 19 |
|
| 20 |
import numpy as np |
| 21 |
import pyvista as pv |
| 22 |
from pyvista import examples |
| 23 |
|
| 24 |
# sphinx_gallery_start_ignore |
| 25 |
# all but first and last image are black in interactive mode |
| 26 |
PYVISTA_GALLERY_FORCE_STATIC_IN_DOCUMENT = True |
| 27 |
# sphinx_gallery_end_ignore |
| 28 |
|
| 29 |
|
| 30 |
# %% |
| 31 |
# Load the example Moon landing image and plot it. |
| 32 |
|
| 33 |
image = examples.download_moonlanding_image() |
| 34 |
print(image.point_data) |
| 35 |
|
| 36 |
# Create a theme that we can reuse when plotting the image |
| 37 |
grey_theme = pv.themes.DocumentTheme() |
| 38 |
grey_theme.cmap = 'gray' |
| 39 |
grey_theme.show_scalar_bar = False |
| 40 |
grey_theme.axes.show = False |
| 41 |
image.plot(theme=grey_theme, cpos='xy', text='Unprocessed Moon Landing Image') |
| 42 |
|
| 43 |
|
| 44 |
# %% |
| 45 |
# Apply FFT to the image |
| 46 |
# ~~~~~~~~~~~~~~~~~~~~~~ |
| 47 |
# FFT will be applied to the active scalars, ``'PNGImage'``, the default |
| 48 |
# scalars name when loading a PNG image. |
| 49 |
# |
| 50 |
# The output from the filter is a complex array stored by the same name unless |
| 51 |
# specified using ``output_scalars_name``. |
| 52 |
|
| 53 |
fft_image = image.fft() |
| 54 |
fft_image.point_data |
| 55 |
|
| 56 |
|
| 57 |
# %% |
| 58 |
# Plot the FFT of the image |
| 59 |
# ~~~~~~~~~~~~~~~~~~~~~~~~~ |
| 60 |
# Plot the absolute value of the FFT of the image. This is a visualization of |
| 61 |
# the **frequency spectrum**, not a spatial image — each pixel represents the |
| 62 |
# amplitude of a frequency component, not a location in the moon landing photo. |
| 63 |
# |
| 64 |
# Note that we are effectively viewing the "frequency" of the data in this |
| 65 |
# image, where the four corners contain the low frequency content of the image, |
| 66 |
# and the middle is the high frequency content of the image. |
| 67 |
|
| 68 |
fft_image.plot( |
| 69 |
scalars=np.abs(fft_image.point_data['PNGImage']), |
| 70 |
cpos='xy', |
| 71 |
theme=grey_theme, |
| 72 |
log_scale=True, |
| 73 |
text='Moon Landing Image FFT Spectrum', |
| 74 |
copy_mesh=True, # don't overwrite scalars when plotting |
| 75 |
) |
| 76 |
|
| 77 |
|
| 78 |
# %% |
| 79 |
# Remove the noise from the ``fft_image`` |
| 80 |
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
| 81 |
# Effectively, we want to remove high frequency (noisy) data from our image. |
| 82 |
# This is still done in the frequency domain — we are modifying the spectrum, |
| 83 |
# not the spatial image. First, let's reshape by the size of the image. |
| 84 |
# |
| 85 |
# Next, perform a low pass filter by removing the middle 80% of the content of |
| 86 |
# the image. Note that the high frequency content is in the middle of the array. |
| 87 |
# |
| 88 |
# .. note:: |
| 89 |
# It is easier and more efficient to use the existing |
| 90 |
# :func:`pyvista.ImageDataFilters.low_pass` filter. This section is here |
| 91 |
# for demonstration purposes. |
| 92 |
|
| 93 |
ratio_to_keep = 0.10 |
| 94 |
|
| 95 |
# modify the fft_image data |
| 96 |
width, height, _ = fft_image.dimensions |
| 97 |
data = fft_image['PNGImage'].reshape(height, width) # note: axes flipped |
| 98 |
data[int(height * ratio_to_keep) : -int(height * ratio_to_keep)] = 0 |
| 99 |
data[:, int(width * ratio_to_keep) : -int(width * ratio_to_keep)] = 0 |
| 100 |
|
| 101 |
# Set an explicit clim with a positive lower bound; zeroed pixels would |
| 102 |
# otherwise break ``log_scale`` (``log10(0) = -inf``) and render as black. |
| 103 |
abs_data = np.abs(data) |
| 104 |
clim = [abs_data[abs_data > 0].min(), abs_data.max()] |
| 105 |
|
| 106 |
fft_image.plot( |
| 107 |
scalars=abs_data, |
| 108 |
cpos='xy', |
| 109 |
theme=grey_theme, |
| 110 |
log_scale=True, |
| 111 |
clim=clim, |
| 112 |
text='Filtered FFT Spectrum (High Frequencies Zeroed)', |
| 113 |
copy_mesh=True, # don't overwrite scalars when plotting |
| 114 |
) |
| 115 |
|
| 116 |
|
| 117 |
# %% |
| 118 |
# Convert to the spatial domain using reverse FFT |
| 119 |
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
| 120 |
# Finally, convert the filtered spectrum back to the spatial domain using the |
| 121 |
# inverse FFT. This is the actual denoised image. |
| 122 |
|
| 123 |
|
| 124 |
rfft = fft_image.rfft() |
| 125 |
rfft['PNGImage'] = np.real(rfft['PNGImage']) |
| 126 |
rfft.plot( |
| 127 |
cpos='xy', |
| 128 |
theme=grey_theme, |
| 129 |
text='Denoised Moon Landing Image (Inverse FFT)', |
| 130 |
) |
| 131 |
# %% |
| 132 |
# .. tags:: filter |