Copyright 2019, Guillaume Duranceau
https://github.com/drnc/rmgradient
rmgradient is a background gradient remover from TIFF files, delivered as a Python script. It builds a background model of the image from input points, and substracts it. This can be used to remove light pollution gradient from astrophotography images for example.
The background model is built from input points coordinates on the image, provided by the user. Those points are assumed to all have the same color once the background is removed. When processing astrophotography pictures, these points are typically chosen in starless dark sky areas of the image.
The background model is built as a multiquadratic radial basis function interpolation of the color values for the chosen points in the image, computed with scipy.
The colors used for the interpolation aren't directly the color values of the input points. A Gaussian blur on the image is applied at the input points coordinates, so as to smooth input data.
Substracting the computed background model from the image would result in a completely dark background on the output, but would also very likely clip information from the input image. Instead, each channel of the background model is darkened so that its minimum value accross the whole image is zero, before being substracted from the input image. This allows to level the background of the input image, while preserving as much information as possible and giving more freedom in post processing.
A few dozen of background points are usually enough to get a quite precise background model on typical astrophotography images.
Note that those points cannot be chosen
too close to the borders of the image.
The convolution kernel size
for the Gaussian blur applied on those points
is 3 times the standard deviation of the Gaussian distribution.
Using the default standard deviation of 8 (--sigma option),
this means that background points should be at least
24 pixels away from the borders.
It's very common that large parts of the image cannot be used as background dark points (like when framing the Milky way or a nebula in an astrophotography). If the background gradient pattern across the image is simple and regular (which is often the case with light pollution), It's possible to help the interpolation process by specifying a series of points on an isoline (which are assumed to have the same background color). The coordinates of those points should be given as negative values:
100 250 <-- background point of coordinate (100, 250) 750 600 <-- another background point 200 400 <-- background point P, starting a series on an isoline -300 -360 | -400 -310 | points (300, 360), (400, 310), (500, 250) -500 -250 |<-- (550, 200) and (600, 140) are assumed to have -550 -200 | the same background value than P (200, 400) -600 -140 | 800 720 <-- back to a standard background point
Note that this can be used to reference points close to image borders.
rmgradient performs all computation with 64 bits floating point values. It writes the resulting TIFF image file with the same type than the input image (if input image is 16 bits TIFF files, output image is a 16 bits TIFF file).
rmgradient can optionally compress the output TIFF file.
The background model can optionally be written in a separate TIFF file.
The standard deviation (sigma) of the Gaussian distribution used to blur the input background points can be specified. As described above, this impacts the minimum distance to the borders of the input points.
Applying the interpolation function to the whole image is a memory consuming operation. To limit memory consumption, rmgradient applies the interpolation by group of limited number of rows (100 by default, but this can lowered to accomodate a user environment with low memory constraints).
The smoothness of the background model approximation can be controlled.
If present in the input image, rmgradient copies the embedded color profile in the output image.
rmgradient.py is a Python 3 script, depending on numpy, scipy and tifffile packages.
To run it and display the usage and options:
python rmgradient.py -h
Usually, coordinates of input points chosen as reference for the background are written in a file:
x1 y1 x2 y2 ...
This file can then used as an input this way:
cat input_points | xargs python rmgradient.py -i image.tif -o output.tif [options] -p
An example image and input points are available in the test data folder, together with a test_sky_image.sh script, to demonstrate how to use rmgradient.