CFA Integer Resample

Purpose and expected result

CFA Integer Resample changes an undebayered Bayer mosaic by an integer factor without mixing the four CFA positions. It runs one filter-aware plan of the image engine: Downsample combines only samples under the same filter, and Upsample replicates each sample within its own filter's lattice. The result is a mosaic of the same arrangement and phase.

The operation modifies the selected still image or replaces its sequence. It is not ordinary interpolation and does not create additional sensor resolution.

At a glance

Item	Current behavior
Menu path	Process > Geometry > CFA Integer Resample
Enabled when	An image or sequence view is active. The target selector lists only views shown as a one-channel mosaic; a target that does not state a two-by-two mosaic is refused when OK is pressed.
Inputs	One undebayered RGGB/BGGR/GRBG/GBRG still or sequence. An image that states no filter array, a colour image and an X-Trans mosaic are refused by name.
Outputs	Same target/sequence at a CFA-preserving integer scale, same sample type, Bayer pattern and phase. No new view/file for stills.
Recommended image state	Calibrated linear CFA data before Debayer.
Preview support	None. Read-only geometry/memory estimates are the only pre-execution feedback.
Sequence support	Supported: every frame is resampled into a new sequence that replaces the view's, with progress and Cancel.
Mask support	Unsupported.
External dependencies	None.

Before you start

1. Verify CFA metadata. The process changes geometry in place; the change is one history step you can undo.
2. Prefer a factor that divides each half-resolution CFA plane: source width and height should be divisible by 2 × factor for the displayed W/factor size to equal the actual downsampled result without discarded edges.
3. Check disk/RAM independently before large Upsample factors. A factor f multiplies pixel count by f²; readouts are capped by 100,000-pixel spin-box limits and can under-report larger computed dimensions.
4. Plan to plate-solve again after resampling if astrometry matters: the result drops the source's solution, because the four filters' samples sit at different places within each block, which one solution cannot describe.

Procedure

1. Open Process > Geometry > CFA Integer Resample and verify Target Image.
2. Choose Downsample or Upsample and set Resample Factor.
3. For Downsample, choose Average, Median, Maximum, or Minimum.
4. Compare Original Px, Target Px, Current Size, and New Size. Calculate the exact result from the formula below when dimensions are odd/not divisible or the readout reaches 100,000.
5. Click OK. Sequence mode shows a progress window; still mode runs without one.
6. Confirm actual dimensions, CFA pattern and history before Debayer.

Complete parameter and action reference

UI label/action	Type and unit	Runtime default	Persisted-default exception	Valid values, step, or constraints	Effect of increasing, decreasing, or selecting	Dependencies and enable/visibility rules	Quality, artifact, runtime, memory, or disk trade-off	Scope: preview/apply, channel, mask, still/sequence	Persistence
Downsample	Radio button	On	None	Mutually exclusive with Upsample	Reduces each filter's lattice by the factor using the selected statistic.	Enables Downsample Mode.	Reduces resolution/data; Average improves random-noise SNR, Median rejects outliers, extrema bias output.	Execute; still/sequence; no mask	Window lifetime; target change/Reset restores On.
Upsample	Radio button	Off	None	Mutually exclusive with Downsample	Replicates each sample within its filter's lattice by the factor on both axes.	Disables Downsample Mode.	Pixel count/memory/disk grow by f²; no new detail is created.	Execute; still/sequence	Window lifetime; target change/Reset restores Off.
Resample Factor	Integer multiplier/divisor	2	None	2–100; step 1; a reduction additionally needs 2 × factor no larger than the even width and height	Larger factors reduce/enlarge more strongly. Every whole factor is accepted and keeps the arrangement's phase, as the tooltip says.	Both directions.	Downsample discards incomplete blocks; Upsample cost grows quadratically and can exceed readout limits.	Execute	Window lifetime; target change/Reset restores 2.
Downsample Mode	Selector	Average	None	Average, Median, Maximum, Minimum	Chooses the statistic over each f×f block of one filter's samples.	Enabled only for Downsample.	Average preserves mean; Median resists outliers; Maximum/Minimum select extremes and bias brightness/noise. Median is costlier.	Downsample only	Window lifetime; target change/Reset restores Average.
Original Px > Width	Read-only pixels	Target width	Live	1–100000 display range	Shows current source width, clamped by widget maximum.	Refreshes on target/image update.	Can under-report sources wider than 100,000.	Readout only	Live.
Original Px > Height	Read-only pixels	Target height	Live	1–100000 display range	Shows current source height.	Same.	Same cap.	Readout only	Live.
Target Px > Width	Read-only pixels	floor(W/2) for default UI formula W/2	Live	1–100000 display range	UI displays integer W/f or W×f. The actual size follows the formula below.	Recalculates with direction/factor.	May differ for odd/non-divisible dimensions or clamp at 100,000.	Readout only	Live.
Target Px > Height	Read-only pixels	floor(H/2) for default	Live	1–100000 display range	UI displays H/f or H×f.	Same.	Same discrepancy/cap.	Readout only	Live.
Current Size	Read-only estimate	Current image/sequence bytes in MB/GB	Live	Two decimals; uses current frame bytes × frame count	Estimates uncompressed image payload, not container overhead.	Sequence count included.	Does not represent working copies or compression.	Readout only	Live.
New Size	Read-only estimate	Computed target width×height×sample bytes×frames	Live	Two decimals	Estimates final uncompressed payload from displayed Target Px.	Recalculates with settings.	Can under-report when dimensions clamp and omits working memory.	Readout only	Live.
Target Image	CFA mono image/sequence selector	Launching view	None	Views shown as a one-channel mosaic	Changes source and resets direction/factor/mode.	Stays enabled during execution; a change does not affect the run already started.	Sequence length drives disk/time.	Execute	Current window only.
OK	Button	Enabled at factor 2 when displayed target differs	None	Valid two-by-two mosaic target	Checks the target, reads the dialog and resamples on a worker as an edit of the still image, or replaces the sequence; records one history step on success.	Disabled while running and when the displayed target equals the current size.	Can allocate very large outputs; still work has no progress/cancel window.	Execute	Window closes after success.
Reset	Button	—	None	Click action	Refreshes source/readouts and restores Downsample, factor 2, Average.	Disabled during execution.	Does not undo an already applied resample; use Image History.	Process window	No preset.
Close	Button	—	None	Click action	Closes without executing pending settings.	Disabled during execution. A window closed during a run stays hidden until it lands.	No change.	Process window	No persistence.

How the algorithm works

Conceptual pipeline

1. Check that the target is a one-channel two-by-two mosaic.
2. If the width or height is odd, crop the last column and/or row first; the crop starts at the top-left pixel, so the phase is unchanged.
3. Run the engine's filter-aware integer plan: Downsample each filter's samples independently with the selected f×f statistic, or Upsample by replicating each sample within its filter's lattice.
4. Publish the still image or replace the sequence, as one history step.

Model and equations

Let source dimensions be W×H, factor f, and per-filter lattice dimensions w=floor(W/2), h=floor(H/2).

Actual Downsample output dimensions are

\[ W_d=2\left\lfloor\frac{w}{f}\right\rfloor,\qquad H_d=2\left\lfloor\frac{h}{f}\right\rfloor. \]

Actual Upsample dimensions are

\[ W_u=2wf,\qquad H_u=2hf. \]

Thus an odd source edge is always lost, and Downsample also drops incomplete f×f blocks of each filter's lattice. The UI's simpler floor(W/f) and W×f readouts agree only for compatible even/divisible dimensions.

For Downsample Average, each output sample of one filter position is

\[ P'(x,y)=\frac{1}{f^2}\sum_{j=0}^{f-1}\sum_{i=0}^{f-1}P(fx+i,fy+j), \]

where P contains only one CFA position. The mean is accumulated in 64-bit float. Median takes the middle value, averaging the two middle values for an even count; Maximum and Minimum take the largest and smallest valid sample. An integer average or median is rounded to the nearest code, ties to even.

Upsample sets P'(fx+i,fy+j)=P(x,y) for all i,j in 0…f-1. Each original 2×2 Bayer cell becomes a 2f×2f region whose four sublattices retain their respective original values. This is not ordinary whole-mosaic nearest neighbor.

Boundary and failure behavior

• The target is checked when OK is pressed. An image that states no filter array is refused with the image states no colour filter array, a colour image with a mosaic has one channel, and an X-Trans mosaic as not a two-by-two arrangement, each under Error running integer resample.
• A reduction whose 2 × factor exceeds the even width or height is refused when the run starts with a CFA reduction factor leaves at least one complete block in every phase. The UI does not preflight this.
• Size readouts clamp to 100,000 pixels and can diverge from actual output. The engine's memory budget refuses an allocation it cannot make, with nothing changed.
• A failure is reported in a message titled Error running integer resample that begins An error occurred when executing the integer resample process: followed by the reason, and in the log; the image and its history are left as they were. If Image History cannot keep the image the resample replaces, the history says so once, in a message titled Image History, this window says nothing more, and nothing changes.
• A sequence another run holds shows the busy warning. Cancel on a sequence's progress window stops the run quietly and leaves nothing in the temporary directory.

Parameter interactions and starting points

Combination	Conservative starting point	Too weak looks like	Too strong looks like	Adjustment order
Factor + divisibility	Factor 2 with W,H divisible by 4	Desired size reduction/enlargement not reached	Edge loss or huge allocation	Check 2f divisibility -> exact formula -> memory -> execute
Downsample Mode + data goal	Average for SNR/mean; Median only for outliers	Noise/outliers remain	Max/Min biases brightness; Median removes legitimate extremes	Choose statistic by goal -> small test -> compare values
Upsample + later Debayer	Lowest factor required	Display remains coarse	Large replicated blocks waste resources without detail	Debayer native baseline -> choose required scale -> upsample once
Metadata + geometry	Solve again after resampling	No solution is available on the result	Assuming the original solution still applies	Resample -> plate-solve

Inputs, outputs, and metadata

• Still execution changes the selected view as one edit, preserves sample type (any of the eight AfterNight stores), Bayer pattern and phase, changes dimensions, and records the history step CFA Integer Resample with the description Binning factor: <f>X - New size: <w>x<h> - Binning method: <mode> or Resample factor: <f>X - New size: <w>x<h>. The size in that text is the dialog's Target Px readout.
• Sequence execution reads every frame of the sequence as recorded, whatever the player has selected, writes the resampled frames into a new sequence of the same container in AfterNight's temporary directory, publishes it only when every frame is written, shows it in the view, and records one history step.
• The result keeps the acquisition record and meaning, drops the header cards and the plate solution, and withdraws any output decision; its lineage names the engine's geometry operation and CFA Integer Resample. A sample every output was written from keeps the source's validity rule; a mask moves with its samples; a sample the step could not write is marked invalid in a mask.
• No auxiliary image/file is produced. Save/export remains a separate action.

Performance

The engine reads each filter's samples in place and writes one result image; a still run holds the source and the result, plus an even-cropped copy when the source is odd. Average, Minimum and Maximum are linear in input pixels; Median does order-statistic work per block. Upsample output/compute/disk scale with f². A sequence writes one new sequence. The displayed New Size omits working memory and may clamp, so use the exact formula and free-space checks for large sequences.

Worked example

• Evidence mode: Source-and-focused-test-derived.
• Input condition/provenance and settings: a 16-bit RGGB mosaic cropped to 23×15, resampled with the defaults (Downsample, factor 2, Average); then a 32-bit float grey image with no filter array.
• Expected checkpoint: the odd edges go (22×14), each lattice of 11×7 reduces to 5×3, and the result is a 10×6 RGGB mosaic whose first red sample is the average of the first four red samples within half a code. The grey image is refused under Error running integer resample naming the missing filter array, and its revision is unchanged.
• Inspect at 100%/statistics/history: verify actual dimensions, Bayer pattern, block/aggregate values and history.
• Limitations: no approved CFA still/sequence was executed through this process.

Warnings and limitations

• This operation replaces the selected view's image or sequence; undo it from the history if needed.
• There is no preview, mask, or channel selector.
• Odd/non-divisible dimensions lose edges and can make Target Px and the history text inaccurate.
• Factors too large for the lattice fail only during execution.
• Upsample adds no resolution and can require f² storage; readouts can clamp at 100,000 and omit working storage.
• The plate solution is dropped; solve the result again if needed.
• Average preserves mean intensity rather than summed flux; Median/Min/Max have different photometric/noise biases.
• Every integer 2–100 is accepted and the plan preserves positional phase; an odd factor is as valid as an even one.

Troubleshooting

Symptom or exact error	Verified cause/validation path	Recovery	Evidence source
Target absent from the selector	View is not shown as a one-channel mosaic.	Open undebayered data with correct Bayer metadata.	Target filter
the image states no colour filter array	The target states no Bayer pattern.	Reopen with correct RAW/CFA metadata.	checkBayerMosaic
a CFA reduction factor leaves at least one complete block in every phase	2 × factor exceeds the even width or height.	Use f ≤ floor(W/2), floor(H/2) and preferably dimensions divisible by 2f.	engine CFA integer plan
Actual size differs from Target Px	UI uses W/f or W×f; the engine crops odd edges and discards incomplete blocks.	Use the exact formulas above; crop to compatible even dimensions first.	UI vs engine plan
New Size looks too small at extreme factor	Target dimension spin boxes cap at 100,000 and the estimate omits working memory.	Calculate externally and reduce factor before executing.	UI ranges/estimate code
No astrometric solution after Apply	The geometry change drops it.	Plate-solve the result.	geometry metadata policy
An Image History error, and nothing changed	Image History could not keep the image the resample replaces; it says so once and the window says nothing more.	Free memory and run again.	GeometryProcessCase.AnEditTheHistoryCannotKeepIsSaidOnceAndChangesNothing
The busy warning	Another run holds the sequence.	Wait for it to land and run again.	runInPlace
Result looks blocky after Upsample	Values are intentionally replicated; no interpolation/detail synthesis occurs.	Debayer at native size or use an appropriate post-debayer resampler for display geometry.	engine integer plan

Related workflow steps

Use Image Calibration before geometry changes. Compare Split CFA and Merge CFA for the positional model, then Debayer afterward. Use ordinary Integer Resample only after Debayer when CFA phase no longer needs preservation.

References and provenance

• AfterNight's image engine and the geometry support unit define all current behavior; no external resampler is invoked.
• The focused window and unit tests cover the reducers, the odd-edge crop, the kept arrangement and the refusals.