Image Statistics

Purpose, audience, prerequisites, and expected result

Image Statistics measures the active image or current sequence frame one channel at a time. Use it to quantify level, spread, clipping, and background noise before or after a processing step. It reads stored pixel values, not the autostretched display.

You need an active view containing a valid mono or RGB image. The result is a read-only table with one Gray column or separate Red, Green, and Blue columns, and an Alpha column for a colour image with alpha. It does not change pixels, metadata, history, masks, or files.

Reachability and availability

Item	Current behavior
Path	Image > Information > Statistics, or Statistics on the active image-view toolbar. The toolbar advertises Ctrl+Shift+T while that view has focus.
Enabled/visible when	A view with an image is active. Opening the dialog disables that view's Statistics actions until the dialog closes; the focused-view shortcut also honors that disabled state. An image load or update in that view enables the actions again, so a second Statistics window can then be opened.
Supported inputs/state	Mono and colour images of every pixel type, signed integers among them. On a sequence view, the table describes the current decoded frame, not all frames. Display mode, autostretch, zoom, channel display, the view's mask, and selections do not limit the sample; samples the image itself marks as holding no data - a registration border, for one - are left out.
Outputs/side effects	A read-only per-channel table. No output image, file, clipboard data, metadata, or history entry is produced.
Persistence	None. Range and Unclipped return to their defaults in each new dialog.
External dependencies	None. Each measurement runs on a worker thread through the engine, in an operation the application's engine session admits; the table changes only when the latest measurement lands.

Before you start

• Select the view you intend to measure. With a sequence, navigate to the frame of interest first.
• Decide whether exact endpoint values are valid signal. With Unclipped off, normalized values at or below 0 and at or above 1 are excluded. Enable it to include finite black/white endpoints - and, for a floating-point image, values below 0 or above 1 - in the reported population.
• Disable autostretch only if you also want to compare appearance. It does not affect the numbers.

Procedure

1. Activate the image or sequence frame and open Image > Information > Statistics. The initial computation runs asynchronously.
2. Confirm that the title names the intended view and that the table shows Gray or Red/Green/Blue columns (with Alpha for a colour image with alpha).
3. Leave Normalized (0.0-1.0) selected for the least ambiguous comparison between pixel types.
4. Compare Count (px) with the expected pixels per channel. The default excludes black/white endpoints. Enable Unclipped to include finite endpoints, then watch the count and percentage change. Non-finite values are excluded in either state.
5. Use Median and MAD for a robust background comparison; use Min, Max, and the count change to diagnose endpoints. Treat BG Noise as a row-difference noise estimate, not a full noise model.
6. Change the numeric range only when you need familiar 8-, 16-, or 32-bit-like units. This rescales the report; it does not convert the image.
7. Close the window when finished. Refresh measures the image again; image load/update notifications refresh the table automatically.

Complete control and action reference

UI label/action	Type and unit	Default	Valid values/constraints	Effect	Enabled, hidden, or ignored rules	Trade-off or side effect	Scope	Persistence
Numeric range selector	Selector; reported level units	Normalized (0.0-1.0)	Normalized (0.0-1.0); 8-bit (0-255); 16-bit (0-65K); 32-bit (0-4000M)	Rescales the already computed table. Normalized values use seven decimal places; other modes use three.	Always available after data exists. It does not change counts or recompute the source image.	A level or a spread is multiplied by the range's full scale and Variance and BWMV by its square, so a variance stays the square of the standard deviation shown beside it. No value is held to the range: a floating image's mean above one shows above full scale.	All reported channels in this dialog	Dialog lifetime only
Unclipped	Checkbox	Off	Off/on	Off retains only samples strictly inside normalized (0,1). On includes every finite sample, including exact black/white endpoints and a floating-point value below 0 or above 1. Either change measures again on a worker. Count and Count (%) expose the retained population.	Applies to all channels; NaN/infinite samples, and samples the image marks as holding no data, are excluded in either mode.	Including endpoints reveals the complete finite population but can let black borders or saturated samples dominate extrema and distribution measures.	Whole current image/frame, all channels	Dialog lifetime only
Count (px)	Read-only integer row	Computed	0 to width x height per channel	Number of finite samples used.	Changes with Unclipped when endpoints exist; unaffected by report range.	A lower count is evidence of excluded endpoints or non-finite values, not automatic proof of processing damage.	Per channel	None
Count (%)	Read-only percent row	Computed	0-100%, seven decimals	100 x retained count / (width x height).	Same filter as Count (px).	Makes endpoint loss comparable across image sizes.	Per channel	None
Mean	Read-only level row	Computed	Selected report range	Arithmetic mean of retained normalized samples.	Recomputed by Unclipped and image updates.	Sensitive to stars, clipping, gradients, and outliers.	Per channel	None
Sigma	Read-only level row	Computed	Non-negative	The population standard deviation, as StdDev.	Recomputed with the retained population.	The same number as StdDev; the row keeps its place in the table.	Per channel	None
Median	Read-only level row	Computed	Selected report range	The exact central retained value; the mean of the two middle values of an even count.	Recomputed with the retained population.	More robust than the mean for sparse stars and outliers, but not a background-only sample.	Per channel	None
Min / Max	Read-only level rows	Computed	Selected report range	Lowest/highest retained finite sample.	Endpoint values disappear when Unclipped is off.	One extreme pixel can set either value. Pair with counts and histogram clipping information.	Per channel	None
StdDev	Read-only level row	Computed	Non-negative	Population standard deviation around the mean.	Uses the retained population.	Sensitive to gradients, stars, and signal structure as well as noise.	Per channel	None
AvgDev	Read-only level row	Computed	Non-negative	Mean absolute deviation from the arithmetic mean.	Uses the retained population.	Robustness is between mean-based spread and median-based MAD.	Per channel	None
BWMV	Read-only robust-spread row	Computed	Non-negative when defined	Biweight midvariance about the median, with a tuning constant of nine MADs; zero when MAD is zero.	Uses the retained population.	Useful against outliers. It is in squared units, scaled by the square of the selected range's full scale.	Per channel	None
Variance	Read-only squared-spread row	Computed	Non-negative	Mean squared deviation from the mean; StdDev^2 apart from rounding.	Uses the retained population.	In squared units, scaled by the square of the selected range's full scale.	Per channel	None
MAD	Read-only robust-spread row	Computed	Non-negative	Median of abs(x - median).	Uses the retained population.	Robust against sparse outliers; it is not multiplied by a normal-distribution consistency factor in this table.	Per channel	None
BG Noise	Read-only level row	Computed	Non-negative; zero when no row has two first differences	Median per-row population standard deviation of first horizontal differences after up to three strict 5-sigma rejection passes, divided by sqrt(2).	Uses the retained population; a left-out sample is skipped and the next retained sample is differenced from the one before it.	Suppresses smooth background level but still reacts to structure and directionality. Compare similar data only.	Per channel	None
Refresh	Push button	Available	Click	Measures the current image or frame again.	Always available.	The table changes when the measurement lands.	This dialog	None
Close	Push button/default dialog action	Available	Click, Enter when default, or close the window	Closes the dialog and re-enables Statistics for its parent view.	Always available.	None.	This dialog	None

Behavior and state model

AfterNight reads each stored pixel component as a fraction x of its pixel type's range before reporting it. An unsigned integer is read over its largest code, a signed integer from its smallest value to its largest - so a signed 16-bit image's -32768 reads as 0 and 32767 as 1 - and a floating-point value as it is. Non-finite floating-point values are discarded, as are samples the image marks as holding no data. With Unclipped off, only 0 < x < 1 contributes. With it on, every remaining finite value contributes.

All rows are computed independently for each stored channel. The current display composition does not merge or hide channels, and a mask does not restrict the measurement. For a sequence, the view supplies the current frame as the image; frame changes that emit an image update trigger a new calculation.

The dialog measures on a worker thread when it opens, on image load/update, when Unclipped changes and when Refresh is pressed; a measurement asked for while another is on its way replaces it, and the table changes only when the latest one lands. A measurement that cannot be made - no engine session, or an image the engine refuses - leaves the table empty. A rename updates the window title only. Changing the numeric range only reformats the last result.

Use these relationships when reading the table:

• mean = sum(x) / N, where N is Count (px).
• variance = sum((x - mean)^2) / N and StdDev = sqrt(variance).
• MAD = median(abs(x - median)).
• BG Noise estimates high-frequency horizontal variation from adjacent differences; it is not the same as global standard deviation.

The table has no sum-of-squares row; the engine's sum and mean of squares reach PixelMath, not this window.

Inputs, outputs, and metadata

The only input is the active view's in-memory image. The tool reads all pixels of the current still or decoded sequence frame and produces transient table cells. It does not write an image, select pixels, alter display transfer functions, append history, change metadata, save a report, or touch disk.

Performance and responsiveness

Work scales with width x height x channels. The median, MAD and biweight rows rank the retained samples exactly: an 8- or 16-bit image, which is what a camera RAW opens as, is ranked from one read of each channel, and a floating-point image reads each channel several times. Every measurement runs on a worker thread, so large frames can take time to settle while the window stays responsive. A measurement may use every thread of the Processing page's limit that is free.

For long sequences, pause playback on a representative frame before opening the dialog. Close the tool when you no longer need continuously updated statistics.

Worked example

• Evidence mode: Source-and-test-derived.
• Starting state, actions, settings, checkpoints, result, and limitations: StatisticsReportTest measures a five-pixel 16-bit row holding 0, 1, 30000, 65534 and 65535. With Unclipped on, Count (px) is 5; with it off the two end codes leave, Count (px) is 3, Min is above 0 and Max below 1. A floating-point row of -0.25, 0, 0.2, 0.4, 1 and 1.5 counts 6 with Unclipped on and 2 with it off. A signed 8-bit row of -128, -127, 0, 126 and 127 with Unclipped off reports Min 1/255 and Median 128/255 in the normalized range. These are synthetic estimator checks, not a claim about a particular astrophotography exposure.

Warnings and limitations

• Statistics describe stored pixel data, not autostretched appearance.
• Unclipped includes exact finite endpoints; with it off, strict endpoint exclusion is active. Neither state performs sigma clipping or highlight/shadow repair.
• The tool measures the full current frame. It has no ROI, mask, or selected- channel sampling mode.
• The bit-range selector is report scaling, not precision conversion.
• No table export or copy action exists. Image Summary has a separate clipboard action but reports metadata, not these statistics.

Troubleshooting

Symptom or exact error	Verified cause/enablement path	Recovery	Evidence source
Statistics is disabled	No active image exists, or this view already owns an open Statistics dialog.	Activate a loaded view, or close the existing dialog for that view.	ActionStateCoordinator.cpp; ImageStatisticsControl.cpp
Count rises when Unclipped is enabled	Finite values <= 0 and >= 1 are restored to the population; non-finite samples remain excluded.	Compare Count (%), Min, Max, and Histogram clipping; leave Unclipped off when endpoints should not influence the report.	ImageStatisticsReport.cpp; ImageStatisticsControl.cpp
Mean and median differ strongly	The distribution is asymmetric or contains bright/dark structure and outliers.	Inspect Histogram Analysis and compare channels before assuming a defect.	ImageStatisticsReport.cpp; StatisticsReportTest.cpp
BG Noise is unexpectedly high	Horizontal structure, gradients, stars, or correlated noise contribute to adjacent-pixel differences.	Compare similar background regions with other analysis tools; do not interpret this full-frame number as a calibrated sensor-noise measurement.	AdvancedStatistics.h; ImageStatisticsReport.cpp
A sequence value changes during playback	Statistics follows the current decoded frame and refreshes on image updates.	Pause on the frame you want to inspect.	ImageStatisticsControl.cpp; ImageViewControl.cpp

Related workflow steps

• Inspect Histogram Analysis next to see where the counted samples lie and whether exact black or white endpoints are populated.
• Use Image Summary for dimensions, pixel type, acquisition metadata, or sequence-level frame/FPS information.
• After saving, use Saving And Exporting and reopen the file before comparing statistics when you need to verify codec precision.
• Continue to Histogram Transformation only after distinguishing measurement from display stretch.

References and provenance

• AfterNight's estimators are the engine's basic and advanced statistics; the window's selection, range and scaling are implemented in ImageStatisticsReport.cpp and its refresh in ImageStatisticsControl.cpp. Test_ControlsStatistics holds them to statistics computed in the suite over every standard fixture.
• StatisticsReportTest.cpp and StatisticsWindowTest.cpp in Test_ControlsStatistics cover every pixel type, grey and colour, endpoint filtering, the image's own validity, background noise, the range scaling, and the window's refresh and latest-result rule; BasicStatisticsTest.cpp and AdvancedStatisticsTest.cpp in Test_EngineImageOps hold the engine's estimators, including identical results at every thread count.
• No external runtime, catalog, network service, or external algorithm package is used by this tool.