17-EFT.WP.Methods.Imaging v1.0 | Chapter 14 Imaging Quality Metrics, SLOs, and Audit

Home ／ Docs-Technical WhitePaper ／ 17-EFT.WP.Methods.Imaging v1.0

Chapter 14 Imaging Quality Metrics, SLOs, and Audit

One-Sentence Goal
Constrain the entire imaging pipeline—from linear radiometric measurement to release decisions—using a reproducible, auditable family of quality metrics SLI.img.* and service-level objectives SLO.*. Coverage spans optics, noise, color, geometry, time/path consistency, arrival time, and runtime performance.

I. Scope & Targets

1. Inputs
   • Data & references: raw or linearized images/sequences y, reference images y_ref, and calibration fixtures for geometry/color/noise.
   • Metadata: unit(x), dim(x), ts, tau_mono, offset/skew/J, exposure and readout modes, optical/filter configurations.
   • Path & arrival time: gamma(ell), n_eff, c_ref, and two-form arrival-time requirements.
2. Outputs
   • Metrics:
     SLI.img.mtf50, SLI.img.mtf_area, SLI.img.psnr, SLI.img.snr, SLI.img.nps,
     SLI.img.prnu, SLI.img.dsnu, SLI.img.deltaE00.P95, SLI.img.wb_error,
     SLI.img.err_geo, SLI.img.dead_pixel_rate, SLI.img.sync_offset, SLI.img.tarr_delta.
   • Decisions: SLO compliance report, violation list with graded remedies; audit_report and manifest.imaging.qc.
3. Boundaries
   • All measurements are taken in the linear radiometric domain; the rendering domain is for visualization only (see Chapter 4).
   • Event cameras, HDR, ToF, and other modalities inherit linearization conventions from their dedicated chapters.

II. Terms & Variables

1. Noise & radiometry
   • sigma_read (read-noise std), k_shot (shot/photon factor), NPS(f) (noise power spectral density).
   • SNR = mu / sigma or SNR_dB = 20 * log10( mu / sigma ), with unit(SNR_dB)="dB".
2. Optics & resolution
   MTF(f), MTF50 (frequency where MTF(f50)=0.5), MTF_area = ( ∫_0^{f_Nyq} MTF(f) df ).
3. Color
   • DeltaE_00 (CIEDE2000), summary stats such as DeltaE_00.P95, DeltaE_00.mean.
   • wb_error = || chroma(meas) - chroma(ref) || at a specified illumination.
4. Geometry
   err_geo = sqrt( ( 1 / N ) * ( ∑ || x_i' - H(x_i) ||^2 ) ), with H the calibration homography/distortion model.
5. Pixel health
   dead_pixel_rate = ( count(bad) / N_pix ), where bad is defined via tol_hot, tol_dead, tol_fpn.
6. Time/path & arrival
   offset, skew, J; arrival-time two forms and delta_form.
7. Error aggregation
   Use high quantiles (Pxx, e.g., P99), sliding windows Delta_t, runtime latency T_proc, and throughput Qps.

III. Axioms P214- (Quality, SLO & Audit Baseline)*

• P214-1 (Linear-domain measurement): compute all quality metrics in the linear radiometric domain; never on gamma/curved renderings.
• P214-2 (Explicit windows & measures): each metric must declare its domain and measure, e.g., ( ∫_Omega • d r ), ( ∫_f • d f ).
• P214-3 (Two-form arrival): metrics involving propagation time must compute both T_arr forms and log delta_form.
• P214-4 (Unified time base): evaluate time-related metrics on tau_mono and publish on ts, carrying offset/skew/J.
• P214-5 (Reproducible & auditable): log algorithms, parameters, ROIs, seeds, and hash_sha256(blob); sign artifacts.
• P214-6 (Units/dimensions): run check_dim(expr) per metric before release and declare unit(x), dim(x).
• P214-7 (Quantile-first decisions): SLOs are judged by high quantiles (P95/P99), not by means.
• P214-8 (Cross-modality parity): for fused modalities, evaluate on a common FOV under a unified A_path(•) (see Chapter 13).

IV. Minimal Equations S214-*

• S214-1 (MSE and PSNR)
• MSE = ( 1 / |Omega| ) * ( ∑_{r ∈ Omega} ( y(r) - y_ref(r) )^2 )
• PSNR = 10 * log10( MAX^2 / MSE )，其中 MAX 为线性域满刻度。
• S214-2 (SNR and NPS)
  SNR = mu / sigma；
  NPS(f) = ( 1 / A ) * E( | F{ n(r) }(f) |^2 ), with F{•} the 2-D Fourier transform and A the area measure.
• S214-3 (MTF and MTF50/area)
• MTF50 = f such that MTF(f) = 0.5
• MTF_area = ( ∫_0^{f_Nyq} MTF(f) d f )。
• S214-4 (PRNU/DSNU indices)
  PRNU = std( flat / mean(flat) )；DSNU = std( dark ) at unified exposure and temperature (see Chapter 8).
• S214-5 (Color difference & white balance)
  DeltaE_00.P95 = P95( { DE2000( Lab_i, Lab_ref_i ) } )；wb_error = || u'v'(meas) - u'v'(ref) ||.
• S214-6 (Geometric error)
  err_geo = sqrt( ( 1 / N ) * ( ∑ || x_i' - H(x_i; k1,k2,...) ||^2 ) ) (H per Chapter 9).
• S214-7 (Pixel health rate)
  dead_pixel_rate = ( count( |gain-1| ≥ tol_gain ∨ |offset_dark| ≥ tol_dark ∨ stuck ) ) / N_pix.
• S214-8 (Sync & arrival consistency)
  S_sync = max( |offset|, |skew|, J )；
  delta_form = | ( 1 / c_ref ) * ( ∫ n_eff d ell ) - ( ∫ ( n_eff / c_ref ) d ell ) |.
• S214-9 (Example composite score)
• Q_img = w_mtf * ( MTF50 / f_Nyq ) + w_psnr * ( PSNR / PSNR_ref )
• + w_color * ( 1 - DeltaE_00.P95 / tol_deltaE )
• + w_geo * ( 1 - err_geo / tol_geo )
• - w_noise * ( NPS_band / NPS_ref )。

Weights w_* are specified in the policy card.

V. Pipeline & Operational Flow M140-*

• Acquisition & linearization: under a fixed RefCond, acquire flat/dark/chart/slanted-edge/pattern; linearize to radiometric space and validate unit, dim.
• ROI & measure declaration: annotate Omega, frequency bandwidth, and windowing; persist roi.json and measure.json.
• Optical metrics: per Chapter 5 (slanted-edge or point-source), compute MTF(f), yielding MTF50, MTF_area.
• Noise metrics: estimate sigma_read, k_shot; compute NPS(f) and band-limited NPS_band.
• Fixed-pattern noise: per Chapter 8, export PRNU, DSNU, and pixel health maps.
• Color metrics: under specified illuminant and color space, compute DeltaE_00 quantiles and wb_error.
• Geometry metrics: with checkerboard/dot-grid, estimate H and err_geo; check extrinsic drift.
• Time/path consistency: evaluate offset/skew/J and both T_arr forms to obtain delta_form.
• SLI aggregation & decision: compose SLI.img.*; check against SLO.* via policy cards; output violation list.
• Audit & signature: produce audit_report; update manifest.imaging.qc; write hash_sha256 and signature.

VI. Contracts & Assertions

• assert mtf50_floor: MTF50 ≥ f_min (units declared as cyc/pixel or lp/mm).
• assert mtf_area_min: MTF_area ≥ A_min.
• assert psnr_min: PSNR ≥ tol_psnr.
• assert snr_min: SNR_dB ≥ tol_snr_db.
• assert nps_band_max: NPS_band ≤ tol_nps_band.
• assert prnu_max: PRNU ≤ tol_prnu; assert dsnu_max: DSNU ≤ tol_dsnu.
• assert color_fidelity: DeltaE_00.P95 ≤ tol_deltaE and wb_error ≤ tol_wb.
• assert geometry_acc: err_geo ≤ tol_geo.
• assert pixel_health: dead_pixel_rate ≤ tol_dead.
• assert sync_ok: |offset| ≤ tol_offset, |skew| ≤ tol_skew, J ≤ tol_jitter.
• assert arrival_consistency: delta_form ≤ tol_Tarr.
• assert audit_reproducible: metric re-measure deviation ≤ tol_repeatability and hashes match.

VII. Implementation Bindings I140-*

• measure_mtf_slanted_edge(img, roi) -> { MTF(f), f50, area }
• estimate_noise_psd(seq, roi) -> { NPS(f), sigma_read, k_shot, NPS_band }
• compute_prnu_dsnu(flats, darks) -> { PRNU, DSNU, map_prnu, map_dsnu, dead_pixel_rate }
• evaluate_color(chart_raw, illum, profile) -> { DeltaE00_stats, wb_error }
• calibrate_geometry(pattern_imgs) -> { H, err_geo, intrinsics, extrinsics }
• eval_sync(meta) -> { offset, skew, J }
• arrival_time_consistency(gamma, n_eff, c_ref) -> { T_arr_const, T_arr_general, delta_form }
• aggregate_sli(metrics) -> SLI_dict
• check_slo(sli, policy) -> { pass, violations[] }
• emit_qc_manifest(sli, policy, hashes) -> manifest.imaging.qc
• export_audit_report(sli, violations, artifacts) -> audit_report

VIII. Cross-References

• Linearization & metrology: Chapter 4; noise & FPN: Chapters 7 and 8; optical resolution: Chapter 5.
• Geometry & registration: Chapter 9; color & WB: Chapter 10; HDR & exposure: Chapter 11.
• Time/path & arrival: Chapter 13; two-form conventions and thresholds: Methods.Cleaning v1.0, Chapter 6.
• SLO & compliant release framework: Methods.Cleaning v1.0, Chapter 10; streaming KPIs: Methods.Cleaning v1.0, Chapter 11.

IX. Quality Metrics & Risk Control

1. Runtime SLIs
   • SLI.svc.proc.P99 = P99( T_proc ) ≤ SLO.svc.proc.P99; SLI.svc.drop ≤ SLO.svc.drop.
   • SLI.img.mtf50.P95 ≥ SLO.img.mtf50.P95; SLI.img.deltaE00.P95 ≤ SLO.img.deltaE00.P95; SLI.img.tarr_delta.P99 ≤ SLO.img.tarr_delta.P99.
2. Risks & playbooks
   • Optical degradation: trigger self-check when MTF drops; suggest refocus or lens service; optionally downscale publication.
   • Noise surge / hot pixels: re-estimate dynamic black; mask pixels; quarantine offending batches on violation.
   • Color drift: switch to backup CCM or re-calibrate; if out-of-bounds, mark degraded color confidence.
   • Sync/arrival anomalies: reset trigger chain and re-evaluate delta_form; hard failures block publication.
3. Audit replay
   Hash, parameters, and versions for all metrics and intermediates (ROIs, edge responses, spectra, etc.) are written to audit_report; paired with TraceID for replay.

Summary
This chapter closes the metrics → SLO → audit loop for the imaging domain, unifying measures, time bases, and path semantics. With contract-driven publication via SLI.img.* and SLO.*, and traceability through manifest.imaging.qc and signatures, imaging quality governance becomes reliable, reproducible, and auditable across modalities, devices, and scenes.