Object SNR Tool

Honest, regional signal-to-noise measurement — pick the structure you care about, pick a background, get a per-channel SNR of that region. No "global image SNR" lies.

1. The two-step workflow

2. Step 1 — Picking a target region

Click Select Target Region. The picker opens on your active image with the auto-background box drawn in gold.

Three selection modes are available from the dropdown at the top:

• Box — click-drag a rectangle. Finalizes in red on release.
• Ellipse — click-drag an ellipse. Same interaction.
• Freehand — click-drag a lasso around any shape. Closes automatically when you release the mouse.

Changing the drawing mode or starting a new drag clears the previous selection — you always have exactly one target region.

The automatic background is a 50×50 pixel source-free box found by scanning the image for the darkest, flattest region a safe distance (100 px by default) from the edges. The box is drawn in gold so you can see where it landed and judge whether it represents a reasonable "nothing-there" background for your target. If it landed somewhere unrepresentative (another faint nebula, a star glow, inside your region of interest), zoom out, click Find Background again (which re-runs the search), or build a more deliberate workflow around the Magnitude tool's four-quadrant verification if you need that level of control over placement.

3. Step 2 — Calculating SNR

Click Calculate SNR. The tool reads the native image (mono stays mono, RGB stays RGB — nothing is converted), masks it with your target region and the auto-placed background box, and reports per channel:

• Stotal(med) — median pixel value inside the target region.
• B(med) — median pixel value in the background box.
• noise(std) — sample standard deviation of the background box's pixels. This is the measured noise level.
• So — net signal above background, clamped at zero: max(Stotal − B, 0).
• SNR — the ratio S_o / noise, as a pure number.
• SNR in dB — the same ratio expressed in decibels, 10 log₁₀(SNR).

For a mono image you get one line labeled L. For an RGB image you get three lines, R, G, B. The results text area is selectable and the Copy Results button sends everything to the clipboard for pasting into a processing log.

3.1 Interpreting the number

The SNR number is a direct answer to the question "how many standard deviations above the sky background does my target sit?" A few reference points:

The dB value doubles for every tenfold increase in linear SNR (power ratio convention): 10× SNR is +10 dB, 100× is +20 dB. The dB axis is more convenient for comparing processing steps on log-scale structures.

4. The three plot tabs

The right-hand panel has a channel selector at the top and three tabs below.

4.1 Distributions

A pair of overlaid histograms showing the pixel-value distribution inside the Background region (blue) and the Object region (orange), normalised to density so they're directly comparable regardless of area. The channel selector picks which plane to plot (L for mono; R, G, or B for color). The bins span from the 0.5th to 99.5th percentile of the combined data, so outliers don't crush the useful range.

What to look for:

• Well-separated peaks — the object histogram sits cleanly to the right of the background. This is the visual confirmation that your SNR number is meaningful.
• Overlapping peaks — the two distributions sit on top of each other. SNR will be very low; the object is at or below the noise floor for this background region.
• Bimodal object histogram — your target region contains both the structure of interest and something else (sky, a star, a brighter neighbour). Redraw tighter.
• Fat-tailed background — your background box isn't clean. Re-roll Find Background, or redraw to avoid stars/faint nebula.

4.2 SNR Chart

A bar chart with per-channel SNR on the left y-axis (linear) and dB on the right y-axis (same bars, different scale). For mono images one bar; for RGB three bars (R, G, B). Fast visual comparison across channels — if your blue SNR sits well below red and green, that's a nudge toward spending more sub-exposure time in blue next session.

4.3 ROI Preview

A cropped view of your target region (expanded by 40 px of context on each side) with the target mask drawn as a translucent red overlay and the background box as a gold tint. Serves one purpose: confirmation that the numbers you're seeing came from where you think they did. The display is autostretched for visibility — it doesn't affect the measurement, which operates on the linear pixels.

5. Choosing a good background

The measured SNR is only as meaningful as the background it's measured against. Three guidelines:

• Representative of your target's local sky. If you're measuring a faint arm on one side of a galaxy, your background should ideally sit on the same side of the frame — gradients across the frame are real, and comparing arm brightness to sky on the far side can bias both your signal (if the gradient varies under the target) and your noise (if the far side has a different texture).
• Source-free. The automatic placement does a decent job finding a dark, flat spot but it's not infallible. If stars or faint nebular wisps leaked into the box, the measured noise will be too high (that extra structure adds to the standard deviation) and your SNR will be artificially suppressed. Click Find Background to re-roll; or if automation fails on a complicated frame, zoom in and visually confirm the gold box sits on actual nothing.
• Big enough. The tool rejects background regions smaller than 25 pixels as too small to characterize noise. The default 50×50 (2500 pixels) is plenty for stable statistics. If you reduce the primary tool's background box size, keep it above ~20×20.

6. Workflow recipes

7. Technical details

The math behind every number the tool reports.

7.1 Signal estimation

The signal estimator is the median of pixel values inside the target mask:

Median is used instead of mean because real target regions almost always contain some contamination — a star within the lasso, a hot pixel residual, a diffraction spike. The mean is pulled by those outliers; the median shrugs them off, so the number you get reflects the typical brightness of your structure, not whatever the brightest hot pixel happened to be. For a well-chosen region on a smoothly-varying nebula or galaxy arm, median and mean agree to within ~1%; for a region containing even a single star the median is dramatically more honest.

The same median estimator is applied to the background region:

and the net signal above background is clamped non-negative:

The clamp exists because a region can legitimately be fainter than the background box (a dust lane in front of a nebula, for example), which would yield a negative S_o and a negative SNR number that doesn't mean anything useful. In that case the tool reports SNR = 0 and the right conclusion is "this region is at or below the measured sky level" — the chart will show it as absent rather than inverted.

7.2 Noise estimation

The noise estimator is the sample standard deviation of the background pixels, using the Bessel-corrected (N-1) denominator:

Standard deviation (not MAD or any other robust estimator) is used here deliberately. In a cleanly-chosen source-free background box, the pixel-value distribution is dominated by Gaussian read-plus-shot noise, and standard deviation is the right estimator of its scale. If stars have leaked in, standard deviation will reflect that — which is correct behavior for an SNR measurement, because those stars are part of what competes with your signal in that part of the frame. A robust MAD estimator would hide the problem and give you an overconfident SNR number; the tool deliberately uses the honest estimator and leaves you to notice (via the distribution plot or the ROI preview) when your background box is dirty.

The Bessel correction (1/(N-1) rather than 1/N) is the unbiased estimator for a sample from a larger population — the usual choice when you're estimating the underlying noise distribution's spread rather than describing the exact pixel population inside the box.

7.3 The SNR definition and the decibel convention

The signal-to-noise ratio is reported as:

This is the ratio of the net signal (median above background) to the noise standard deviation — essentially, "how many σ above the sky is my structure?" It's the same quantity underlying the Rose criterion of visibility (SNR ≥ 5), underlying detection significance claims in astronomy ("3σ detection" means SNR = 3), and underlying every statement about faint-source photometric uncertainty.

The decibel version uses the power-ratio convention:

This gives +10 dB per factor-of-10 improvement in linear SNR. Equivalently:

which inverts the dB value back to a plain ratio. Some image-quality tools use the amplitude-ratio convention 20 log₁₀(SNR) instead; that convention doubles every dB number compared to this tool, so if you're comparing SNR_dB from another tool, verify which convention it uses before concluding anything about relative quality.

7.4 Why SNR varies spatially (and why this tool is honest about that)

The underlying truth every pixel's SNR obeys is:

Both the signal and the noise are properties of a region, not of the whole frame. In a deep-sky frame the quantities take on wildly different values depending on where you look:

• Background sky: I(x,y) ≈ B, so SNR ≈ 0. The sky background is, by definition, where SNR is zero.
• Faint outer halo of a galaxy: I(x,y) sits a fraction above B, SNR ∈ [1, 10].
• Spiral arm: I(x,y) sits several σ above B, SNR ∈ [10, 50].
• Galaxy core: I(x,y) ≫ B, SNR in the hundreds or thousands.
• Star: typically SNR in the thousands or ten-thousands, approaching the detector's full-well regime.

A single-number "global SNR" that collapses this to one value will always be dominated by whichever brightness regime occupies the largest number of pixels in the frame — usually the background, sometimes a bright extended object. The number it produces tracks how bright your frame is on average, which is not what you want to know; you want to know how well-detected the structure you care about is.

By asking you to pick the region and the background, this tool keeps both terms of the SNR definition under your control. The number it returns is the real-and-specific SNR of the thing you asked about, measured against a specific reference. That's the only kind of SNR number that means anything.

7.5 Per-channel handling

Each channel is treated as an independent image. A mono image yields one SNR number labeled L; an RGB image yields three independent SNR numbers labeled R, G, B. The tool does not combine channels into a single "colour SNR" (no such concept exists cleanly — color SNR depends on luminance weighting choices and on perceptual-vs-scientific intent). You get three honest numbers, one per channel, and compare them yourself.

For narrowband palettes where each channel is a different line (SHO, HOO, etc.), per-channel SNR is directly interpretable as the detection quality of each emission line and feeds straight into integration-time planning.

7.6 Stacking and the √N scaling

For reference, in the Poisson/Gaussian noise regime that governs deep-sky stacks, SNR grows with the square root of the number of co-added subs:

So doubling your integration gives a √2 ≈ 1.41× SNR improvement (+1.5 dB); quadrupling it doubles the SNR (+3 dB). If you measure SNR on two stacks of the same object with the same cropping and your observed ratio falls meaningfully short of this prediction, something is wrong with the longer stack — bad calibration, drifting guiding, poor rejection of outliers in integration, or a gradient contaminating one of the nights. This tool is the right instrument to diagnose that.