Files
aimsight-labeling/docs/archive/knowledge/quality-metrics.md
T

18 lines
3.0 KiB
Markdown
Raw Normal View History

# The image-quality metrics of a target
Six columns of the "advanced" export, per sensor and per measurement: `peak`, `flux`, `sharpness`, `fwhm`, `dominance`, `snr`. Computed by the GPU of recent cameras (`aimsight_optical` 6.7.11) on the sensor window in the DIFF image, in units of that image. Filled on LENZSPITZE and NADELHORN (pumping of October 2026), empty elsewhere; for an older camera, `~/optical_emulator` recomputes them in `quality/` (`run.md`).
Source: `~/optical_emulator/docs/cuda_phase_correlation.md`, §14 (definitions of the reference code). "Disc": the neighborhood of the peak, of radius `min(H, W)/6` (at least 3 px); "background": all the rest of the window (`bg_mean`, `bg_sigma`).
| Metric | Definition | What it measures | Reading |
|---|---|---|---|
| `peak` | `max(I)` | the brightness of the return | high, saturation; low, underexposure or missed flash. All targets dropping: the lighting (LED, flash); a single one: that target (fouling, misalignment) |
| `flux` | flux of the disc minus the background | the light of the target, background removed | reflectivity of the target (fouling, aging, flash), not ambient light; drops well before `error_value` exceeds its threshold |
| `sharpness` | `Σ(I²) / (ΣI)²` | concentration of the energy (higher = sharper) | drops: drifting focus, thermal turbulence, degrading target, i.e. more measurement noise. Independent of the gain, but not of the background (not removed): to follow over time on one target, not to compare between cameras |
| `fwhm` | width at half height around the peak, background removed [px] | width of the spot | rises: defocus, turbulence, motion blur. Truncated to the disc: under strong defocus, it flattens instead of rising |
| `dominance` | `(second peak − background) / (peak − background)`, from 0 (clean) toward 1 (ambiguous) | does a single target dominate the window? | high: a comparable second reflection, a stray bright point, or the target has disappeared and something else is the brightest: the correlation can latch onto the wrong pattern, which `error_value` alone does not see. A single hot pixel outside the disc is enough to make it rise |
| `snr` | `(peak − background) / bg_sigma` | the signal above the background noise | low: marginal return, even when `peak` looks normal. Independent of exposure and gain. The emulator marks `empty` below 8 |
In short: `peak` and `flux` give the brightness, `sharpness` and `fwhm` the sharpness, `dominance` and `snr` the validity. Together, they say **why** a measurement degrades, where `error_value` only says that it degrades.
These are qualities of the image of **one target**: they say nothing about the health of the reference network (the non-negotiable rules, at the end of `labeling/knowledge/displacement-columns.md`; `AIMsight_agentic_support_project_context.md`, §5.5 and §5.6). The readings in the table are those of the source; what they are worth on our phenomena (rain, condensation, obstruction, sun) is confirmed in the data (`labeling/knowledge/measurement-chain.md`).