Skip to content

Patch PCA Denoise

Purpose and expected result

Patch PCA Denoise reduces noise with a local, patch-based principal-component filter. At regularly sampled reference positions it compares every patch in a square lookup neighborhood with the reference patch, estimates a local second-moment matrix, removes eigen-components below an automatically estimated noise threshold, reconstructs all neighboring patches, and overlap-averages the results.

The expected result is a same-size, same-type still or replacement sequence with weaker random noise. Repeated local structure can help define strong components, but the current implementation does not rank or reject patches by similarity: Lookup Size includes every position in the window. Strong thresholds can remove faint stars or texture; large neighborhoods can mix unrelated structure; and sparse sampling can leave uneven or even uncovered output regions.

At a glance

Item Current behavior
Menu path Process > Noise Reduction > Patch PCA Denoise Filter
Enabled when An image or sequence is active, the same process is not already attached to it, and no attached process holds exclusive access
Inputs One grayscale or RGB still or sequence; a still CFA mosaic is treated as one grayscale sample plane unless debayered first; an optional active mask must match target geometry
Outputs Destructively replaces a still or, after successful temporary processing, a sequence; Take snapshot creates an auxiliary still; dimensions, channels, and source pixel type are retained unless a selected component reconstructs RGB; no file is written directly
Recommended image state Usually a finite, normalized linear master after calibration, registration, stacking, gradient correction, and color calibration but before stretching
Preview support Asynchronous RT preview with the same captured parameters, channel conversion, and mask blend as Apply; selected-area preview and display-only autostretch are available
Sequence support Supported through the generic temporary replacement-sequence path with per-frame processing, progress, and cooperative cancellation; configured CFA auto-debayer can produce RGB frames
Mask support Supported after denoising for preview, still Apply, and sequence Apply; white applies, black protects, gray blends, and Invert Mask reverses weights
External dependencies Bundled G'MIC 3.5.4 command denoise_patchpca, run through AfterNight's optional G'MIC adapter; no separate model, catalog, network service, or installation. In a build without G'MIC, Preview and Apply are disabled and their tooltip says why
Non-finite samples An image holding a NaN or infinite sample is refused before G'MIC sees it, and left unchanged

Before you start

  • Finish calibration, registration, stacking, gradient correction, and color calibration first. The lookup window uses every local patch, so walking noise, registration defects, gradients, hot pixels, and compression texture enter the local model rather than being identified as outliers.
  • Use finite samples normalized to [0,1]. An image holding a NaN or infinite sample - such as the empty border of a registered frame - is refused with "the image holds a sample that is not a finite number, which G'MIC cannot process" and left unchanged. Integer inputs are converted into a float32 G'MIC buffer as fractions of full scale; float inputs are processed in place and double inputs are reduced to float32 precision. Output is always clamped to [0,1] before conversion back to the original type, so negative or above-one float data is clipped even at Strength 0.
  • This process needs G'MIC, which AfterNight normally includes. In a build without it, Preview and Apply are disabled and their tooltip reads "This build of AfterNight carries no G'MIC library, so this process cannot run."
  • Treat the live defaults as 4 / 7 / 11 / 7, not the dormant builder values 1.8 / 7 / 8 / 7 or G'MIC's omitted-argument defaults. AfterNight explicitly supplies all four values and its constructor calls initialize() after the controls are connected.
  • Tune on a representative Preview area containing background, fine stars, and a real edge. Patch Size and Lookup Size can increase work by orders of magnitude, while lower Spatial Sampling values make reference centers denser.
  • Duplicate the target or verify Image History. Apply replaces a still only after success; a sequence is processed into temporary storage and committed only after every frame succeeds.

Procedure

  1. Select a finite normalized master and open Process > Noise Reduction > Patch PCA Denoise Filter. Confirm Target Image, select RGB or one deliberate component, and leave Preview selected.
  2. Begin with live defaults: Strength 4, Patch Size 7, Lookup Size 11, and Spatial Sampling 7. Use a Preview area before running the full frame.
  3. Judge at Zoom 100% under one Display Mode. Toggle Switch image and inspect smooth background, faint stars, edges, and each color channel.
  4. Tune Strength first. Lower it if faint structure disappears; raise it only while residual random noise falls without flat or blocky patches.
  5. Change Patch Size or Lookup Size one at a time. Larger values make the model less local and dramatically more expensive. Reduce Spatial Sampling only when denser estimation materially improves coverage or consistency.
  6. Verify the full-frame preview, mask state, target, and Channel before selecting Apply. For a sequence, monitor frame progress; Cancel preserves the original and removes temporary output.

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
Strength Double value plus slider, dimensionless noise multiplier 4.00 Builder starts at 1.80, but the live process constructor immediately sets 4.00; G'MIC omitted-argument default 1.8 is never used 0.00..16.00; two decimals; spin step 0.01; slider has 1000 positions across the range Sets the cutoff Strength × estimated noise; higher values retain fewer leading PCA components, while 0 retains all returned components Always enabled while idle; noise is estimated automatically and is not exposed Higher values smooth more and can erase faint structure; the slider resolves about 0.016, coarser than the spin step Preview/Apply; selected channel/component; still/sequence; before mask Captured per execution; Reset and successful Apply return 4.00
Patch Size Integer value plus slider, pixels per side 7 None 2..21, step 1; even sizes use asymmetric left/right extents Sets each N×N patch and PCA dimension N²; larger patches model broader structure Always enabled while idle; no validation against image dimensions Covariance storage/work grows with N⁴ and eigendecomposition approximately with N⁶; large values can mix detail and become extremely slow Preview/Apply; selected channel/component; still/sequence Captured per execution; Reset and successful Apply return 7
Lookup Size Integer value plus slider, patch positions per side 11 Builder starts at 8, but the live process constructor immediately sets 11 2..21, step 1; all L² offsets are used, including the reference region Expands or contracts the local window used to build the matrix and reconstruct neighboring patches; it does not select only similar patches Always enabled while idle; no distance threshold or group-size control Work grows approximately with L²; larger windows mix less-local structure and increase aggregate writes Preview/Apply; selected channel/component; still/sequence Captured per execution; Reset and successful Apply return 11
Spatial Sampling Integer value plus slider, reference-center stride in x and y 7 G'MIC omitted-argument default is 5, but AfterNight explicitly supplies 7 1..16, step 1; a center runs only where both coordinates are divisible by this value Higher values skip more centers and run faster; lower values produce denser estimates Always enabled while idle; not automatically limited by Patch/Lookup Size Approximate center count falls as 1/S²; overly sparse support can produce uneven weights or uncovered black pixels, especially when S > Patch Size + Lookup Size - 1 or at the far image edges Preview/Apply; selected channel/component; still/sequence Captured per execution; Reset and successful Apply return 7
Target Image Image selector Attached/active target None Open eligible image or sequence Changes source, Channel list, Preview, active mask, and Apply destination Active image required; duplicate/exclusive gates apply Wrong target is destructively replaced after success Preview/Apply source and active mask Target lifecycle; Reset does not change it
Channel Combo box RGB for RGB; disabled Grayscale for mono None RGB, Red, Green, Blue, CIE L* (Lightness), CIE a, CIE b, CIE c, CIE h, CIE X, CIE Y (Luminance), CIE Z, HSL Hue, HSL Saturation, HSL Lightness, HSV Hue, HSV Saturation, HSV Value, YCbCr Luma, YCbCr Blue Chroma, YCbCr Red Chroma RGB sends all stored channels to G'MIC, which filters each channel independently with one shared noise threshold; a stored/component choice filters one plane and reconstructs/copies the others Component choices require RGB; mono selector disabled Independent RGB decisions can change color around weak detail; component conversions add buffers and may clip or shift color Preview/Apply; active mask post-blend; still/sequence Captured by current selection; Reset does not restore RGB
Preview Checkbox Checked None Checked/unchecked Enables/disables asynchronous full-image or selected-area execution Process idle Runs the complete G'MIC filter; repeated large-patch previews can be very expensive Preview only; mask-aware; current still/current displayed sequence frame Window lifetime; Reset does not change it
Preview pane Image viewer Current result None Pan with scroll bars/dragging; toolbar actions below Shows the mask-aware result without modifying the target Exists only with Preview Holds source/result and display buffers in addition to G'MIC workspaces Preview only Window lifetime
Fit To Window Preview-toolbar action Preview fitted when first displayed None Click or Ctrl+Shift+F Fits the whole image uniformly in the viewport Preview image required Good overview; resampling can hide sparse-grid or patch artifacts Preview display only Viewer lifetime
Zoom In Preview-toolbar action Current scale None Click; 1.25× relative zoom within viewer limits Enlarges around viewport center Preview image required Improves local inspection; more panning Preview display only Viewer lifetime
Zoom Out Preview-toolbar action Current scale None Click; 0.8× relative zoom within viewer limits Reduces around viewport center Preview image required More context, less fine detail Preview display only Viewer lifetime
Zoom 100% Preview-toolbar action Current scale None Click; absolute scale 1.0 Shows the viewer's one-pixel scale Preview image required Preferred for judging residual noise, black gaps, patch texture, and stars Preview display only Viewer lifetime
View Channel Preview-toolbar combo box RGB for RGB; disabled Grayscale for mono None RGB, Red, Green, Blue, CIE L* (Lightness), CIE a, CIE b, CIE a=R / b=G (Chrominance), CIE c, CIE h, CIE X, CIE Y (Luminance), CIE Z, HSL Hue, HSL Saturation, HSL Lightness, HSV Hue, HSV Saturation, HSV Value, HSI Hue, HSI Saturation, HSI Intensity, YCbCr Luma, YCbCr Blue Chroma, YCbCr Red Chroma Selects only the displayed stored/derived component Separate from process Channel; RGB target required for components Can reveal channel artifacts without changing pixels Preview display only Viewer lifetime; not in process history
Display Mode Preview-toolbar combo box Inherits target Linear or Autostretch when preview is created None Linear, Autostretch Changes display only; Autostretch uses cached full-target source statistics Autostretch enables settings action Aggressive display stretch can exaggerate residual noise/artifacts Preview display only Preview-view state; not in process history
Autostretch Settings Preview-toolbar dialog action Disabled in Linear mode; inherits target-view settings Shared-view values load when opened Opens Target Background, Shadows Clipping, linked channels, neutralization/color calibration, saturation/background, brightness, contrast, banner, Reset, and Close; Apply Stretch is hidden Tunes preview display only Autostretch mode only; disabled while dialog open Can change visual judgment without changing denoise Preview display only Viewer-scoped; not in Patch PCA parameters
Switch image Preview-toolbar action Denoised preview shown None Click; Space; right-click; Ctrl+Shift+O Toggles result/original Preview required Comparison only; no recomputation Preview display; still/current frame Preview-window lifetime
Preview area Preview-toolbar checkbox Unchecked None Checked/unchecked; move/resize overlay Processes a selected area and composites it over the original preview Preview required Safest way to reduce tuning cost; full Apply remains full-frame Preview and Take snapshot; mask cropped consistently Preview-window lifetime; Reset unchanged
Take snapshot Preview-toolbar action No snapshot None Click or Ctrl+Shift+K Re-executes settings and opens <target> - Patch PCA Denoise - SNAPSHOT Preview required Adds a complete execution and auxiliary image; display stretch is not baked in New still; target unchanged; area/mask honored Each action creates a separate image view
Apply Button/action Enabled None One action Captures four parameters and selected Channel, then processes the full still or every sequence frame Target must lock; controls disabled during commit Full-frame local matrices and eigensolves dominate; component conversion, mask, sequence encode, and temporary storage add cost Destructive still/sequence replacement after success Records partial history, then calls Reset; Target, Channel, Preview, and preview toolbar remain
Reset Button/action Enabled while idle None One action Restores Strength 4, Patch Size 7, Lookup Size 11, and Spatial Sampling 7 Does not reset Target, Channel, Preview, mask, or preview toolbar Checked Preview schedules another full execution Configuration only Immediate; also invoked after successful Apply
Close Button/action Enabled while idle None One action Closes without applying current settings Disabled during Apply Active preview cancellation is requested and reaches the running command through G'MIC's abort flag; pending result discarded Window/execution No process settings persist after destruction
Sequence progress / Cancel Shared per-process/frame progress and Cancel Created for sequence Apply None Cooperative Cancel through G'MIC's abort flag and between frames; frame/task text updates during temporary processing; G'MIC reports no progress within a frame Sequence target only Decode/encode, temporary disk, and serial per-frame Patch PCA work add cost Sequence Apply; captured parameters, Channel, and mask apply to every frame Cancel removes temporary output; success replaces sequence, records one history step, and resets parameters

How the algorithm works

Float bridge and automatic noise estimate

AfterNight runs the command through its G'MIC adapter, one G'MIC session per run, and hands G'MIC 3.5.4 the planar samples as single-precision fractions of full scale: a tightly packed float image in place, anything else - integer or double input, or padded rows - as a working copy. Before the call it checks that every sample is finite and refuses the image otherwise. The command's answer is held to [0,1] and converted back to the image's type, rounded to nearest and held to the type's range. The command estimates one noise scale before processing channels:

\[ \hat\sigma=\frac{1.4826}{\sqrt{20}} \operatorname{median}\left( \left|\Delta I-\operatorname{median}(\Delta I)\right| \right), \]

where ΔI is G'MIC's 2-D Laplacian. For an RGB image the image statistics cover all stored channel samples, so the same scalar threshold is used while each channel is filtered independently. The process exposes no manual noise value. Real edges and undersampled stars can influence this robust estimate.

Lookup neighborhood and PCA cutoff

Let Patch Size be N, Lookup Size be L, Spatial Sampling be S, and d=N². A reference center is evaluated only when both x mod S and y mod S are zero. Its reference patch is the vector X. For every one of the L² lookup offsets, the neighboring patch Y_i contributes

\[ M=\frac{1}{L^2}\sum_{i=1}^{L^2}(Y_i-X)(Y_i-X)^T. \]

This is an uncentered second-moment matrix relative to X, not a conventional mean-centered covariance and not a similarity-selected patch group. The code eigendecomposes M, takes lambda_j=sqrt(abs(mu_j)), and retains the leading returned components while

\[ \lambda_j \ge \text{Strength}\,\hat\sigma. \]

All remaining coefficients are set exactly to zero. Each lookup patch is reconstructed as

\[ \hat Y_i=X+QKQ^T(Y_i-X), \]

where K is one for retained components and zero otherwise. Strength 0 keeps every returned component, but the float bridge, overlap arithmetic, output clamp, and type conversion can still change pixels.

Aggregation, color, and boundaries

Every reconstructed lookup patch is drawn into a full-frame accumulation image through a Gaussian patch mask. A separate weight image receives the same mask; the final output divides accumulated samples by max(weight,0.01). Patch reads outside the image use Neumann/replicated boundaries, while writes outside the image are clipped.

For RGB, the G'MIC formula runs in parallel over pixels and channels. It builds a separate N² × N² matrix and cutoff reconstruction for each channel; there is no joint color covariance. A shared noise threshold does not prevent different components being retained in R, G, and B, so weak colored structure can shift. A selected CIE/HSL/HSV/YCbCr component is converted, filtered as one plane, and reconstructed through the shared Channel path. A derived component's reconstruction saturates an integer image at its type's range and keeps a floating image's values outside [0, 1].

The weight floor does not restore source pixels. If reference centers and their lookup patches do not cover a location, its accumulated value remains zero and the output can be black there. Interior support from one center spans Patch Size + Lookup Size - 1 pixels per axis; therefore Spatial Sampling above that span guarantees interior gaps, and right/bottom edge remainders can be uncovered sooner. The live 7/11/7 geometry covers the ordinary interior, but extreme allowed combinations require a full-frame check.

Parameter interactions and starting points

  • Strength × estimated noise is the only shrinkage threshold. A noisy or edge-dominated estimate changes every local cutoff.
  • Patch Size changes both spatial support and matrix dimension. Going from 7 to 14 doubles patch width but makes the dense matrix 16 times larger and raises cubic eigensolver work by roughly 64 times before lookup work.
  • Lookup Size does not make a similarity search more selective. Increasing it includes every additional local offset and can blend unrelated features.
  • Spatial Sampling is a two-axis stride. Moving from 7 to 1 can create about 49 times as many reference centers away from boundaries.
  • Patch/Lookup/Sampling coverage must be checked together. Fast sparse settings can trade consistent overlap for grid texture or zero-weight gaps.
  • Channel changes the data model. RGB is three independent filters with one shared noise estimate; a derived component adds conversion/reconstruction and may affect color even though only one component was selected.

Inputs, outputs, and metadata

  • Output dimensions and source pixel type are retained. G'MIC works in float32, final samples are clamped to [0,1], and integer output is rounded.
  • The result keeps the source image's header cards and metadata.
  • The process does not update linear/nonlinear processing-state provenance. Do not infer state only from the fact that denoising was applied.
  • Still Image History records Strength, Patch Size, Lookup Size, Spatial Sampling, and, for a colour image, in channel: <channel>.
  • A still CFA mosaic is treated as one stored sample plane unless debayered first. For a sequence, shared preprocessing may auto-debayer configured CFA frames and the replacement sequence can therefore be RGB.
  • Sequence frames use one captured parameter set, are processed to temporary storage, and replace the original sequence only after success. Cancel/error removes the temporary output.
  • Take snapshot creates a still from the current preview scope. It does not change the target or bake the preview's display autostretch into pixel values.

Performance

For image width W, height H, channels C, patch side N, lookup side L, and sampling stride S, the approximate number of reference centers is WH/S² per channel. At each center the dominant work is about O(L²N⁴ + N⁶): L² outer products and projections of N²-element vectors, plus a dense eigendecomposition of the N² × N² matrix. These are scaling guides, not measured timings. G'MIC's * formula evaluates in parallel, so CPU core count and memory bandwidth also matter.

G'MIC holds the float input/result, a full-frame per-channel aggregation image, a full-frame weight image, and per-thread patch, matrix, eigensystem, and projection vectors. AfterNight also holds destination, preview, component, and mask buffers as needed. Sequence mode adds decode/encode buffers and temporary disk roughly comparable to an output sequence. No approved astronomical image was timed and peak memory was not instrumented; use a Preview area and avoid changing multiple size parameters upward together.

Worked example

These source-and-calculator examples explain thresholds and scaling. They were not executed on an approved astronomical dataset and make no quality claim.

Suppose the Laplacian estimator returns sigma=0.020. At live Strength 4, the component cutoff is

cutoff = 4 × 0.020 = 0.080

If the leading returned component standard deviations are 0.120, 0.070, and 0.030, only the first remains. Strength 2 lowers the cutoff to 0.040 and retains the first two; Strength 0 retains all returned components.

For default geometry, the matrix dimension is 7²=49, it contains 49²=2401 values, and each sampled center examines 11²=121 lookup patches. Changing only Patch Size to 14 makes the dimension 196 and matrix contain 38,416 values. Changing only Spatial Sampling from 7 to 1 can make reference centers about 49 times denser. These ratios omit parallel overhead, boundaries, previews, and I/O.

Warnings and limitations

  • The UI builder's starting values 1.8/7/8/7 are not the live process's: the constructor, Reset, and post-Apply values are 4/7/11/7.
  • Lookup Size is labeled as a similar-patch search window, but the bundled command uses every one of its L² positions; it performs no distance ranking or rejection.
  • Shrinkage is a hard cutoff. Strong values can remove faint stars, thin filaments, and low-contrast texture abruptly or create flat/patchy regions.
  • RGB channels are filtered independently with one shared estimated noise value. Weak color detail can shift, and derived-component round trips can alter color and, in an integer image, clip at its type's range (a floating image keeps values outside [0, 1]).
  • Spatial Sampling can create grid inconsistency or black zero-weight gaps for sparse parameter combinations and far edges. Always inspect a full-frame result at 100% after tuning on an area.
  • Even patch and lookup sizes have asymmetric integer extents. Oversized windows are accepted even when larger than the image and rely on replicated boundary samples.
  • Float/double values outside [0,1] are clamped on output; double input is processed at float32 precision; integer output is requantized.
  • An image holding a NaN or infinite sample is refused by name and left unchanged; repair or crop non-finite samples first.
  • Without G'MIC in the build, the process cannot run: Preview and Apply stay disabled.
  • The Windows and Linux builds give slightly different results for the same image and settings: G'MIC's script weighs its patches with the platform's own mathematical library, and a last-bit difference there can change how many components a patch keeps. Each build gives the same result every time; do not compare a Windows result with a Linux one sample by sample.
  • Apply automatically resets all four algorithm parameters after success. Target, Channel, Preview, mask, and preview-toolbar state remain.
  • Cancellation is cooperative through G'MIC's abort flag. A costly internal matrix/eigen operation may delay the visible response; sequence Cancel is also checked between frames.

Troubleshooting

Symptom Likely cause Check or correction Evidence
Dialog starts at 4/7/11/7, not 1.8/7/8/7 Live constructor overwrites builder values Use the live values; Reset returns them too Process constructor and initialize()
Large black spots or a grid appear Spatial Sampling leaves zero-weight pixels Reduce Spatial Sampling; enlarge support carefully; inspect the full frame at 100% Vendored aggregation and weight floor
Fine stars disappear abruptly Strength threshold retains too few PCA components Lower Strength and compare under the same stretch Vendored cutoff loop
Increasing Lookup Size changes structure rather than finding better matches Every lookup position is included Reduce Lookup Size around mixed structures; it is not a similarity-ranked group Vendored lookup loops
Runtime jumps after a modest Patch Size change Matrix dimension is Patch Size squared and eigensolve is dense Return toward 7; tune on Preview area Vendored matrix/eigen path
Colors shift Per-channel PCA decisions or a component conversion round trip Compare RGB and a deliberate luminance component on a duplicate; inspect View Channel G'MIC channel loop and shared Channel path
Negative highlights or values above one are clipped The G'MIC result is held to [0,1] Normalize a working copy before denoising GmicProcessSupport.cpp
Error: the image holds a sample that is not a finite number The image has a NaN or infinite sample, which G'MIC cannot process Repair or crop the non-finite samples, then retry; the image was left unchanged GmicProcessSupport.cpp
Preview and Apply are disabled; tooltip says the build carries no G'MIC library This build was made without G'MIC Use a build that includes G'MIC, or another denoiser GmicProcessSupport.cpp
Preview differs from Apply Preview area, mask, target/channel, or display stretch changed Match all controls and use one Display Mode; Apply always processes the full target Shared preview/process path
Cancel appears delayed G'MIC has not reached an abort check inside costly work Wait for the current operation; use smaller patches/lookup or a smaller Preview area next time G'MIC adapter abort flag

References and provenance

  • G'MIC denoise_patchpca reference lists the command's four arguments and upstream omitted-argument defaults. AfterNight bundles G'MIC 3.5.4 and explicitly supplies all arguments, so the vendored gmic_stdlib.gmic remains the behavioral source of truth.
  • G'MIC mathematical expressions defines image median ic, matrix eig(), and the parallel * evaluation used by the bundled command.
  • G'MIC PCA reference describes the general eigenvector projection/reconstruction convention. Patch PCA Denoise uses its own local hard-cutoff implementation rather than invoking that command.
  • G'MIC project and JOSS citation identifies the upstream framework, authorship, and primary project publication.