Drizzle

Variable-pixel linear reconstruction, and more commonly referred to as Drizzle, was developed for Hubble Space Telescope (HST) work by Andy Fruchter and Richard Hook [FruchterHook1997], initially for the purposes of combining dithered images of the Hubble Deep Field North (HDF-N). This algorithm can be thought of as a continuous set of linear functions that vary smoothly between the optimum linear combination technique (interlacing) and shift-and-add. This often allows an improvement in resolution and a reduction in correlated noise, compared with images produced by only using shift-and-add.

There is an excellent page summarising the technique and providing a good graphical representation of how the pixel data "drizzles" down from the coarse input pixel grid onto a finer output pixel grid here.

The degree to which the algorithm departs from interlacing and moves towards shift-and-add depends upon how well the PSF is subsampled by the shifts in the input images. In practice, the behavior of the Drizzle algorithm is controlled through the use of a parameter called the pixel fraction, that represents the amount by which input pixels are shrunk before being mapped onto the output image plane. At a pixel fraction of 0, drizzle is equivalent to pure interlacing; at a pixel fraction of 1, it is equivalent to shift-and-add.

To understand the difference between drizzle and the interpolation methods of applying registration, firstly consider how the standard interpolation method works. The registration data takes the form of a 3x3 Homography matrix, which encodes an 8 degrees of freedom linear transformation from one set of coordinates to another (i.e. from each frame to the reference frame). This is used to map the values from each pixel in each input image onto the correct place in the output image, aligning the output with the reference image. The actual alignment uses an interpolation method, which can be selected in the registration options. Interpolation results in a smearing of the point spread function, especially when upscaling images. It can also result in artefacts, although Siril implements a clamping mechanism to minimize this.

Drizzle, by contrast, turns each pixel in the input image into a droplet, and drizzles it through a grid pattern onto the output reference frame. Each droplet has a size, and by choosing a scaled-up output pixel grid but smaller droplet size you can achieve improved resolution if your imaging train is undersampled. (If your sampling is correct for the resolving power of your telescope, Drizzle can't produce detail beyond the diffraction limit.) This comes at the cost of increased image noise: because each droplet "paints" a smaller area in the output image, the average coverage of droplets per output pixel in the final stack is reduced.

Note that Drizzle doesn't replace all of the registration process: you still use the Global Star Alignment, 1-2-3 Star Alignment, Comet registration or whichever registration method you wish prior to drizzling: it is an alternative only to the interpolation method used when applying registration.

Uses of Drizzle

There are 3 main reasons you may prefer Drizzle to using an interpolation method to apply registration.

  • Resolution enhancement If your image is significantly undersampled, you may be able to achieve an improvement in resolution using Drizzle that you could not achieve using the Interpolation upscaling x2 registration option.

  • CFA imaging If your images have a CFA pattern (i.e. if you use a one-shot camera or dSLR), Drizzle provides a significant improvement over debayering. This is sometimes referred to separately to Drizzle as Bayer Drizzle, but really it's exactly the same process. When drizzling a CFA image the CFA color of the current droplet sets which channel of the output image it lands on, whereas when drizzling a mono image all droplets land in the (only) mono output channel. Drizzling CFA images avoids the artefacts that occur with all debayering algorithms, which gives improved noise characteristics when strongly stretching images. This supports improved noise reduction and deconvolution for drizzled CFA sequences compared with debayered and registered CFA images, and makes the shadows look better.

  • Avoid artefacts It is possible to drizzle a sequence using scale = 1, pixel fraction = 1 and achieve essentially the same result as applying registration with one of the interpolated methods. You might wish to consider this if you see interpolation artefacts from the standard interpolated method (though these are generally effectively suppressed by the clamping feature). Note that drizzle can produce different artefacts of its own (see the "Some Common Issues" heading below) however these can be completely avoided by choice of drizzle kernel or by having a greater number of input frames, and are generally handled perfectly by stacking.

Limitations of Drizzle

  • Drizzle is a little slower compared to interpolation, particularly the preferred square kernel. If you are using older or slower hardware, you may prefer the legacy method.

  • When used for upsampling, Drizzle achieves its resolution improvements at the cost of increased image noise. Therefore you may wish to collect more integration time when drizzling than when using an interpolation-based upsampling method.

  • The above issue is especially true for CFA images. Consider that only 1 in 2 of the pixels are green, and only 1 in 4 are red or blue. Therefore for the red or blue channels, CFA drizzle effectively already involves the same level of reduction in droplet coverage as a 2x upscale drizzle. If you upscale on top of that, you need as much droplet coverage as you would for a 4x upscale drizzle! Therefore it is generally recommended to drizzle CFA images at scale = 1.

Comparison

The following image shows a comparison between drizzle and the legacy upscaling method. The image is Ha extracted from an OSC session with a dual band filter. On the left you can see the result of the legacy OSC_Extract_HaOIII script, which extracts the Ha data captured by the red pixels in the OSC Bayer matrix as a half-size image and uses OpenCV upscaling with lanczos4 interpolation to produce an image tat matches the size of the OIII image.

On the right you can see the result of the updated OSC_Extract_HaOIII_drizzle script (available through the siril-scripts repository), which extracts the Ha data captured by the red pixels in the OSC Bayer matrix as a half-size image and drizzles it using scale = 2.0, pixel fraction = 0.5, to produce an image that matches the size of the OIII image.

Viewing it at 100% scale it is clear that the drizzled stack restores much of the resolution of the optical system that is undersampled by the spaced-out red pixels in the Bayer matrix: it looks much sharper, and the numbers confirm it: the average fwhm in the left-hand image is 3.59, whereas in the right-hand image it is 3.25.

Comparison between interpolation upscaling and drizzle

Comparison between interpolation upscaling and drizzle

Workflow and User Interface

Mono Workflow

Calibration

For mono images nothing changes in the calibration tab. Calibrate as you normally would.

Registration

Drizzle settings

Registration tab showing drizzle settings

The Drizzle workflow begins with registering your sequence. This is done in the usual way but you must ensure that the chosen registration algorithm does not apply the registration but only saves the registration data in the sequence. For the 1-2-3 stars, two-pass global star alignment, image pattern alignment, KOMBAT and comet alignment methods nothing extra needs to be configured; for the global star alignment method, the Save transformation in seq file only checkbox must be active, otherwise the registration will be applied using interpolation.

Click Go register to register your sequence.

Drizzling

Once the sequence has registration data you are now ready to drizzle it. This time, select the Apply Existing Registration registration method. This will reveal a Drizzle checkbox. Checking this checkbox will show the Drizzle-specific options.

Scale

Scale sets the scale of the drizzle output image with regard to the input image. A typical drizzle scale for an undersampled mono image is 2.0. This means that the input will be drizzled onto an output pixel grid with twice the resolution. (If your input reference image was 1024 x 512 pixels, your output image would be 2048 x 1024 pixels.) Note: becase the image represents the same area of the sky, although there are twice as many pixels along each axis in the output image, effectively each output pixel is half as wide and half as tall.

Tip

The greater the scale, the sparser each drizzled output image and the fewer pixels end up getting stacked into each output pixel. This results in a noisier image: the resolution gain provided by drizzle comes at the expense of noise. This has to be mitigated by using a greater overall integration time than you would need without drizzling to a greater resolution.

Pixel Fraction

Pixel Fraction sets the size of the droplet taken from the input grid. Consider a drizzle scale of 2.0: since the output pixels are half as wide and half as tall, that means that in order for each input pixel "droplet" to be the same size as an output pixel it should be shrunk to half the linear dimensions. This is a pixel fraction of 0.5. A good rule of thumb is that the pixel fraction should be roughly the reciprocal of the drizzle scale (with some kernels it helps to set it a little bigger than this, in order to reduce pixels that receive zero input from any drizzled droplets.

There is scope for experimentation with the pixel fraction: setting a larger pixel fraction means that each input droplet will influence more output pixels. On the other hand, setting a smaller pixel fraction means that each input droplet will influence fewer output pixels. The "point" kernel is a special case where the pixel fraction is zero (and with this kernel selected, the pixel fraction setting has no effect).

Droplet Model

Siril's Drizzle implementation provides several droplet models:

  • Square. This models the droplet as a square droplet aligned exactly with the input pixel. It is accurately mapped to the output reference frame. This and the Turbo method are the only flux preserving drizzle kernels, so these should be chosen if there is a desire to use the output for accurate photometry.

    Tip

    Accurate photometry is important for the PCC and SPCC color calibration processes, therefore if these are to be used later in the workflow it is strongly recommended to choose a drizzle kernel that preserves flux.

  • Point. This models the droplet as a point at the center of the input pixel. It is mapped to the output reference frame and only ever influences the output pixel on which it lands.

  • Turbo. This is a simplification of the Square kernel. It assumes that rotation between the input and output reference is negligible. This results in a much faster computation, but is approximate. It is a "quick and dirty" kernel originally intended for use within the HST workflow where it was used to generate input for an initial stack that was used for pixel rejection and then discarded, with the square kernel being used for the final drizzle. You may find it usable for some purposes, especially where scale = pixfrac = 1.0, but treat it with caution.

  • Gaussian. This models the droplet as a Gaussian centered on the center of the input pixel. This may provide improved resolution recovery, and will limit the pixel fraction to ensure all output pixels receive some droplet coverage, but it is not flux preserving.

  • Lanczos2 and Lanczos3. These kernels model the droplet as a Lanczos function centered on the center of the input pixel. As with the Gaussian kernel, these may provide improved resolution recovery, but are not flux preserving. You may wish to experiment with drizzle kernels to find the one that provides the best-looking results with your data.

Initial Pixel Weighting

When a droplet lands on the output pixel grid, it may cover more than one output pixel. In fact, one output pixel may be covered by multiple droplets, by only a fraction of a droplet or even by no droplets at all. The contribution of each input pixel can be weighted by the master flat, so that pixels from areas with higher SNR (less vignetting) are weighted more highly. Unless you have peculiar flats this makes only a very small difference.

To enable the master flat, check the Include master flat in initial pixel weighting checkbox.

Warning

The master flat must be set in the Calibration tab!

Let's Drizzle!

Once all the options are set, click the Go register button again.

Stacking

You can now stack your drizzled sequence as normal. Note that for some combinations of drizzle scale and droplet size, some rejection models will work better than others. In particular, if you have significant numbers of "zero input" or null pixels, there will be fewer values to use in rejection. MAD may be a good one to try if your usual rejection method struggles.

The GIF below shows a comparison of a stack of 37 images, in one case with registration applied using interpolation and in the other case with registration applied using drizzle. It is clear that the stack made with drizzled data is significantly sharper than the one using data registered using interpolation.

Crop of two stacked images, showing a significant improvement in sharpness in the drizzled stack.

Comparing registration applied with drizzle and with interpolation. Click to enlarge view.

CFA Workflow

CFA Calibration

For one-shot color (OSC) images, uncheck the Debayer before saving checkbox. This represents a change to previous workflows, but for drizzling it is essential that the CFA pattern is preserved in the drizzle input sequence.

CFA Registration

The Drizzle workflow begins with registering your sequence. As described above, you must ensure that the chosen registration algorithm does not apply the registration but only saves the registration data in the sequence.

Tip

OSC users: you will notice that the red notification text will now identify that a supported CFA pattern has been detected.

Click Go register to register your sequence.

CFA Drizzling

Once the sequence has registration data you are now ready to drizzle it. This time, select the Apply Existing Registration registration method. This will reveal a Drizzle checkbox. Checking this checkbox will show the Drizzle-specific options.

CFA Scale

Scale sets the scale of the drizzle output image with regard to the input image. In OSC camera images each pixel only records a single color: red, green or blue. The pixels have a color filter array (CFA) applied to them and this determines which pixels respond to red, green and blue wavelengths. Thus all the pixels are sparsely distributed compared with a mono sensor in which all pixels are sensitive to whatever light passes through the filter. In both Bayer pattern and X-Trans CFAs the red and blue pixels are particularly sparse in the input frames, therefore increasing the drizzle scale above 1.0 will require even more frames to provide enough drizzle coverage and reach an acceptable level of noise.

For a typical OSC sensor application where the seeing is well matched to the nominal sampling of the sensor it is recommended to apply CFA drizzle with scale = 1.0 and pixfrac = 1.0. This will restore resolution in each color channel (which is effectively being undersampled because of the spacing of the colored pixels in the CFA) and avoid conventional debayering artefacts. If you wish to upsample the image as well by using scale > 1.0, be aware that the pixels available in each channel will be getting even more sparse, and you will need even more data to ensure adequate coverage and contain noise to an acceptable level.

Tip

For OSC drizzle, start with scale = pixel fraction = 1.0.

CFA Pixel Fraction

Pixel Fraction sets the size of the droplet taken from the input grid. The same comments apply here as are described above for the mono workflow.

CFA Droplet Model

The same choice of drizzle kernels applies for CFA drizzle as for mono drizzle. Note that the kernels that are particularly prone to generating null pixels can be tricky when used for CFA drizzle. If you have tens of thousands of frames as in a planetary video, turbo may work fine (and will be fast!) however for deep sky sequences with smaller numbers of frames it is recommended to stick to the square or Gaussian kernels (and bear in mind as mentioned above that Gaussian is not flux preserving, so if you intend to do anything involving photometric techniques square is preferred).

CFA Initial Pixel Weighting

As with mono drizzling a master flat may be specified. To enable the master flat, check the Include master flat in initial pixel weighting checkbox.

Warning

The master flat must be set in the Calibration tab!

Let's Bayer Drizzle!

Once all the options are set, click the Go register button again.

Stacking your CFA data

You can now stack your drizzled sequence as normal, noting the same comments on rejection as for mono drizzle (these may be more apparent with CFA drizzle if you have inadequate coverage to support some of the outlier rejection algorithms, owing to the greater sparseness of input pixels in each channel).

Tip

If you are drizzling your CFA data to gain resolution, it is possible you may be disappointed when comparing results with stacked debayered images. There are generally gains, but they may be marginal (e.g. a few percent improvement in fwhm) and generally will not be nearly as impressive as the resolution gains to be had from drizzling undersampled mono data.

The reason for this is that debayering already restores some of the lost resolution. The various debayering algorithms work differently but they generally all rely on principles of spatial and spectral correlation to infer some of the resolution missing in one channel based on information obtained from the other channels. [Losson2010]

The real reason to drizzle CFA data is that the drizzled result has much cleaner noise. It looks less "grainy" (i.e. it lacks the structure that can be seen in the background of a typical debayered CFA stack) and is thus easier to reduce using noise reduction techniques and gives more consistent data for photometric applications such as color calibration. When stretched hard to bring out faint features just above the background, the resulting background looks more natural.

Bayer Drizzle Comparison

The animation below shows a comparison between CFA drizzle with two different pixel fractions and two of the classical debayering algorithms.

Showing a comparison of VNG, RCD and CFA drizzle with 2 different pixel fractions

Comparison of CFA drizzle (here captioned as Bayer Drizzle) with classical debayering algorithms

  • VNG is provided as a basic reference: note the color artefact around the brighter stars.

  • RCD is quite good with round objects like stars.

  • Bayer Drizzle 1.0 gives results very close to RCD but with a better noise and background

  • Bayer Drizzle 0.5 gives better resolution at the cost of more noise. The trade-off that pixel fraction gives between resolution and noise is evident. With a smaller pixel fraction CFA drizzle needs more data to achieve the same noise performance.

Some Common Issues

Tip

DON'T PANIC - the following results may look a bit weird when you view an individual drizzled sub, but they are not bugs - the algorithm is functioning as intended. In most cases they naturally resolve themselves during stacking, in the remaining cases they can be resolved by changing the drizzle parameters or by including more frames of data.

Moiré Patterns

Due to the nature of the drizzle algorithm, when upscaling some output pixels may not receive any input. These are referred to as "null pixels" and they have a zero value. Some kernels compensate for this, effectively by limiting the pixel fraction, so that all output pixels receive some input, but others do not.

Output pixels that don't receive any input are black: since they typically occur in patterns based on the geometry of the transformation from the input frame, they typically look like Moiré patterns, as shown below:

Showing patterns of zero-weighted pixels with the turbo kernel

Showing patterns that result from null pixels in a drizzled image

Don't worry about this! Siril ignores pixels that are exactly 0 in stacking, so as long as you have enough input frames and the dither positions are suitably scattered, all the pixels will receive coverage from enough pixels and the output stack will be fine. However if you are stacking with a lower number of input frames and this is causing problems, try a different drizzle kernel. Here is exactly the same image drizzled with exactly the same scale and pixel fraction, but with the square kernel instead of the turbo kernel. The result is different, and the patterns are no longer evident.

Using the square kernel produces an image with no strange patterns

Using a different drizzle kernel can eliminate patterns of null pixels

Patchy Stacks

One issue you may see when stacking drizzled data, if there are too many null pixels, is that you may get an odd patchy appearance in the final result:

Showing a patchy result from stacking drizzled data with too many null pixels

Typical patchy appearance of a stack of drizzled data with too many null pixels / not enough frames

This typically occurs with the point, turbo or lanczos kernels. You can fix it by using the square or Gaussian kernels or by having more input frames.

Commands

Siril command line

seqapplyreg sequencename { -upscale | -drizzle { [-scale=] [-pixfrac=] [-kernel=] [-flat=] } } [-interp=] [-noclamp] [-layer=] [-framing=] [-prefix=] [-filter-fwhm=value[%|k]] [-filter-wfwhm=value[%|k]] [-filter-round=value[%|k]] [-filter-bkg=value[%|k]] [-filter-nbstars=value[%|k]] [-filter-quality=value[%|k]] [-filter-incl[uded]]
Applies geometric transforms on images of the sequence given in argument so that they may be superimposed on the reference image, using registration data previously computed (see REGISTER).

The output sequence name starts with the prefix "r_" unless otherwise specified with -prefix= option.

The option -upscale activates interpolated x2 upscaling of the images created in the transformed sequence.

The option -drizzle activates the DRIZZLE alogrithm, which can take the additional options: -scale= sets the image scale factor (default = 1.0); -pixfrac= sets the pixel fraction (default = 1.0). The -kernel= argument sets the DRIZZLE kernel and must be followed by one of point, turbo, square, gaussian, lanczos2 or lanczos3. The default is square. The -flat= argument specifies a master flat to weight the drizzled input pixels (default is no flat). The -ocseq argument specifies generation of an output_counts sequence with the additional prefix "oc_".

The pixel interpolation method (when not using DRIZZLE) can be specified with the -interp= argument followed by one of the methods in the list no[ne], ne[arest], cu[bic], la[nczos4], li[near], ar[ea]}. If none is passed, the transformation is forced to shift and a pixel-wise shift is applied to each image without any interpolation.
Clamping of the bicubic and lanczos4 interpolation methods is the default, to avoid artefacts, but can be disabled with the -noclamp argument.

The registration is done on the first layer for which data exists for RGB images unless specified by -layer= option (0, 1 or 2 for R, G and B respectively).

Automatic framing of the output sequence can be specified using -framing= keyword followed by one of the methods in the list { current | min | max | cog } :
-framing=max (bounding box) will project each image and compute its shift wrt. reference image. The resulting sequence can then be stacked using option -maximize of STACK command which will create the full image encompassing all images of the sequence.
-framing=min (common area) crops each image to the area it has in common with all images of the sequence.
-framing=cog determines the best framing position as the center of gravity (cog) of all the images.

Filtering out images:
Images to be registered can be selected based on some filters, like those selected or with best FWHM, with some of the -filter-* options.


Links: register, stack

Note that the introduction of true drizzle has necessitated some changes to existing command arguments for clarity. The old -drizzle argument to the register and seqapplyreg commands (which used to activate x2 upscaling using interpolation) has been renamed -upscale.

seqapplyreg has a new argument -drizzle which, together with some related arguments, activates true drizzle.

It is not possible to activate true drizzle from the register command, so you need to do two steps:

  • First, use register {sequence} -2pass to generate registration data in the sequence;

  • Second, use seqapplyreg {sequence} -drizzle to drizzle the sequence.

References

[Losson2010]

Olivier Losson, Ludovic Macaire, Yanqin Yang. Comparison of color demosaicing methods. Advances in Imaging and Electron Physics, 2010, 162, pp.173-265, section 2.2.2. https://hal.science/hal-00683233/document

[FruchterHook1997]

A. S. Fruchter and R. N. Hook. (1997) A novel image reconstruction method applied to deep Hubble Space Telescope images. Proc. S.P.I.E. vol. 3164. https://arxiv.org/abs/astro-ph/9708242