Image Sensitivity (IS): Metrics and Methods

• Image Sensitivity (IS) is a family of response measures that map perturbations in energy, space, physics, or semantics to quantifiable changes in imaging outputs.
• It encompasses diverse applications including detector calibration, sub-pixel characterization, inverse problem analysis, and perceptual quality assessment across various imaging systems.
• Practical insights involve methods such as calibration constants, spatial sensitivity maps, Jacobian evaluations, and variance analysis to optimize imaging performance and reduce error.

Searching arXiv for papers on “image sensitivity” and closely related usages across sensor calibration, imaging systems, and model analysis. {} Image Sensitivity (IS) denotes different but structurally related quantities across imaging science. In detector calibration it can mean the ratio of measured signal to deposited energy, $\chi(E)=S/E_{\rm inc}$; in sub-pixel characterization it denotes the spatial variation of a pixel’s responsivity within its $2$-D area; in coherent or astronomical imaging it is the minimum signal power or faintest flux that yields a specified SNR; in inverse problems it appears as a Jacobian or “Sensitivity Field”; and in machine learning and image quality assessment it measures the response of a classifier or perceptual metric to controlled perturbations (Failor et al., 2012, Mahato et al., 2018, Khachaturian et al., 2021, Wright et al., 2010, Barbastathis, 2024, Denton et al., 2019). This suggests that IS is best understood as a family of response measures linking image formation or image interpretation to perturbations in energy, space, physics, or semantics.

1. Distinct formal meanings of Image Sensitivity

Across the cited literature, the term is attached to different observables, different perturbation models, and different output spaces. Some uses are tied to hardware calibration, some to forward models of wave propagation, and some to black-box predictive systems.

Domain	Formal object	Representative usage
XUV phosphor imaging plates	$\chi(E)=S/E_{\rm inc}$	Counts per joule or counts/pJ (Failor et al., 2012)
Intra-pixel characterization	$S(x,y)$	Local sensitivity within one pixel (Mahato et al., 2018)
Coherent photonic receivers	$\mathrm{IS}=\mathrm{NEP}$	Minimum signal optical power at SNR unity in a 1 Hz output bandwidth (Khachaturian et al., 2021)
Astronomical instrumentation	Sensitivity limit at specified SNR	Faintest flux or surface-brightness detectable in a given exposure (Wright et al., 2010)
Inverse scattering and phase imaging	$S(\mathbf r;\theta)=\partial\psi/\partial\theta$	Gradient of scattered field with respect to object parameters (Barbastathis, 2024)
Classifier auditing and IQA	Counterfactual, Sobol, or perceptual sensitivity	Output change under semantic or distortion perturbations (Denton et al., 2019, Bahr et al., 23 Jun 2025, Hepburn et al., 2023)

These definitions are not interchangeable. In one setting IS is a calibration constant; in another it is a spatial map, a detection threshold, a Jacobian, or a variance decomposition. The common structure is a mapping from a perturbation or excitation to a measurable change in signal, estimate, or decision.

2. Detector responsivity, noise-equivalent sensitivity, and detection limits

In phosphor-based imaging plates for XUV imaging, image sensitivity is the conversion efficiency between deposited XUV energy and scanned digital counts. The defining equations are

$$S=\chi(E)\,E_{\rm inc}, \qquad E_{\rm inc}=N_{\rm ph}h\nu,$$

so $\chi(E)$ has units of counts per joule and is usually quoted in counts/pJ over $60\,{\rm eV}<E<900\,{\rm eV}$. Calibration at ALS beamline 6.3.2 yielded an average $\chi(E)\simeq 25\pm 15$ counts/pJ over $2$0–$2$1 eV, with representative absolute-calibration values of $2$2 counts/pJ at $2$3 eV, $2$4 counts/pJ at $2$5 eV, $2$6 counts/pJ at $2$7 eV, $2$8 counts/pJ at $2$9 eV, and $\chi(E)=S/E_{\rm inc}$0 counts/pJ at $\chi(E)=S/E_{\rm inc}$1 eV. Small modulations near the low-energy region and near the C K and O K edges reflect overcoat absorption, and quantitative use requires background subtraction, flat-fielding, and fading control (Failor et al., 2012).

In coherent silicon-photonic imaging, IS is defined as the minimum signal optical power that yields a signal-to-noise ratio of unity within a $\chi(E)=S/E_{\rm inc}$2 Hz output bandwidth, equivalently the noise-equivalent power. The formulation is tied to heterodyne I/Q detection with a $\chi(E)=S/E_{\rm inc}$3 hybrid:

$\chi(E)=S/E_{\rm inc}$4

Carrier suppression reduces the LO-induced DC term, and balanced I/Q detection suppresses common-mode LO fluctuations. The proof-of-concept $\chi(E)=S/E_{\rm inc}$5 IQ sensor array demonstrates $\chi(E)=S/E_{\rm inc}$6 resolution over range accuracy and free-space FMCW ranging with $\chi(E)=S/E_{\rm inc}$7 resolution at $\chi(E)=S/E_{\rm inc}$8 m distance, while the row-column read-out architecture requires only $\chi(E)=S/E_{\rm inc}$9 interconnects for $S(x,y)$0 sensors (Khachaturian et al., 2021).

For nondestructive sensors such as the Skipper CCD, sensitivity is optimized by varying the number of reads per pixel. If each read has rms noise $S(x,y)$1, then

$S(x,y)$2

The framework supports static ROI assignment and dynamic “energy-of-interest” triggering. A demonstration with $S(x,y)$3 gives $S(x,y)$4 in-ROI versus $S(x,y)$5 out-ROI, and a Poisson-based allocation on LDSS-3 spectrograph data gives a total readout time of about $S(x,y)$6 of the “all $S(x,y)$7” baseline (Chierchie et al., 2021).

In astronomical instrumentation, image sensitivity is the faintest flux or surface-brightness detectable at a specified SNR under a stated exposure time, spatial sampling, and spectral resolution. For IRIS on TMT, the governing expression is

$S(x,y)$8

With $S(x,y)$9 hours total integration, the imager reaches SNR $\mathrm{IS}=\mathrm{NEP}$0 at $\mathrm{IS}=\mathrm{NEP}$1, $\mathrm{IS}=\mathrm{NEP}$2, $\mathrm{IS}=\mathrm{NEP}$3, and $\mathrm{IS}=\mathrm{NEP}$4. At $\mathrm{IS}=\mathrm{NEP}$5, the finest-scale IFS reaches SNR per spectral channel $\mathrm{IS}=\mathrm{NEP}$6 at $\mathrm{IS}=\mathrm{NEP}$7, $\mathrm{IS}=\mathrm{NEP}$8, $\mathrm{IS}=\mathrm{NEP}$9, and $S(\mathbf r;\theta)=\partial\psi/\partial\theta$0 in the same integration time (Wright et al., 2010).

A related but domain-specific detection-limit usage appears in the FAST H I imaging of M51. Starting from an rms of $S(\mathbf r;\theta)=\partial\psi/\partial\theta$1 mJy beam$S(\mathbf r;\theta)=\partial\psi/\partial\theta$2 per $S(\mathbf r;\theta)=\partial\psi/\partial\theta$3 km s$S(\mathbf r;\theta)=\partial\psi/\partial\theta$4 channel and the optically thin relation

$S(\mathbf r;\theta)=\partial\psi/\partial\theta$5

the study quotes a practical $S(\mathbf r;\theta)=\partial\psi/\partial\theta$6 H I column-density sensitivity of $S(\mathbf r;\theta)=\partial\psi/\partial\theta$7 (Yu et al., 2023).

3. Intra-pixel sensitivity variation and sub-pixel calibration

In scientific CMOS characterization, image sensitivity often refers to intra-pixel sensitivity variation (IPSV), also called intra-pixel response non-uniformity (IRNU): the spatial variation of a pixel’s responsivity to incident light within its $S(\mathbf r;\theta)=\partial\psi/\partial\theta$8-D area. The motivation is precision photometry and astrometry, where even sub-pixel shifts of a point-spread function can modulate integrated flux if the pixel is not spatially uniform. A front-side illuminated CMOS sensor with pixel pitch $S(\mathbf r;\theta)=\partial\psi/\partial\theta$9 was probed using a $$S=\chi(E)\,E_{\rm inc}, \qquad E_{\rm inc}=N_{\rm ph}h\nu,$$0 W incandescent lamp, a $$S=\chi(E)\,E_{\rm inc}, \qquad E_{\rm inc}=N_{\rm ph}h\nu,$$1 pinhole, a Carl Zeiss GF-Planachromat $$S=\chi(E)\,E_{\rm inc}, \qquad E_{\rm inc}=N_{\rm ph}h\nu,$$2 objective with $$S=\chi(E)\,E_{\rm inc}, \qquad E_{\rm inc}=N_{\rm ph}h\nu,$$3, and a three-axis stage of $$S=\chi(E)\,E_{\rm inc}, \qquad E_{\rm inc}=N_{\rm ph}h\nu,$$4 accuracy. For $$S=\chi(E)\,E_{\rm inc}, \qquad E_{\rm inc}=N_{\rm ph}h\nu,$$5 nm and $$S=\chi(E)\,E_{\rm inc}, \qquad E_{\rm inc}=N_{\rm ph}h\nu,$$6, the Airy first-zero radius is $$S=\chi(E)\,E_{\rm inc}, \qquad E_{\rm inc}=N_{\rm ph}h\nu,$$7. A raster over a $$S=\chi(E)\,E_{\rm inc}, \qquad E_{\rm inc}=N_{\rm ph}h\nu,$$8 pixel area with $$S=\chi(E)\,E_{\rm inc}, \qquad E_{\rm inc}=N_{\rm ph}h\nu,$$9 produced $\chi(E)$0 spot positions, and $\chi(E)$1 sub-frames were averaged at each position (Mahato et al., 2018).

The forward-modeling formulation writes the local response as

$\chi(E)$2

or, with an Airy spot centered at $\chi(E)$3,

$\chi(E)$4

One implementation subdivides the pixel into a $\chi(E)$5 grid of $\chi(E)$6 sub-pixels and fits the $\chi(E)$7 unknown $\chi(E)$8 by weighted least squares with uncertainties $\chi(E)$9 and nonlinear Levenberg–Marquardt optimization, without explicit smoothness regularization beyond the Airy-spot convolution physics (Mahato et al., 2018). A related study discretizes the response on an $60\,{\rm eV}<E<900\,{\rm eV}$0 grid, uses AIC/BIC to select $60\,{\rm eV}<E<900\,{\rm eV}$1, and fits a weighted chi-square with

$60\,{\rm eV}<E<900\,{\rm eV}$2

where $60\,{\rm eV}<E<900\,{\rm eV}$3 and $60\,{\rm eV}<E<900\,{\rm eV}$4. That study compares the forward fit to Wiener deconvolution and reports that Wiener filtering smooths edges and loses low-amplitude, high-frequency central structure (Mahato et al., 2018).

Quantitatively, the $60\,{\rm eV}<E<900\,{\rm eV}$5 IPSV map yields normalized sensitivities with mean $60\,{\rm eV}<E<900\,{\rm eV}$6, maximum $60\,{\rm eV}<E<900\,{\rm eV}$7, minimum $60\,{\rm eV}<E<900\,{\rm eV}$8, and RMS variation $60\,{\rm eV}<E<900\,{\rm eV}$9, together with a systematic drop-off toward the pixel edges and highest sensitivity near the geometric center. The reported two-dimensional map shows a radially symmetric fall-off, attributed to partial vignetting by the pixel microlens and front-side circuitry, with smaller asymmetries correlating with guard-rings and metal traces. The $\chi(E)\simeq 25\pm 15$0 forward fit reports edge sub-pixels as low as $\chi(E)\simeq 25\pm 15$1, central sub-pixels up to $\chi(E)\simeq 25\pm 15$2, and overall $\chi(E)\simeq 25\pm 15$3–$\chi(E)\simeq 25\pm 15$4 variation; a hypothesis F-test with $\chi(E)\simeq 25\pm 15$5 rejects flat sensitivity even in the central region. Incorporating the extracted IPSV map as a calibration filter in the data pipeline is reported to reduce photometric errors due to sub-pixel motion to below $\chi(E)\simeq 25\pm 15$6 ppm for typical jitter amplitudes in one study, while the other estimates photometric errors at the $\chi(E)\simeq 25\pm 15$7–$\chi(E)\simeq 25\pm 15$8 level for a typical $\chi(E)\simeq 25\pm 15$9 motion if no correction is applied (Mahato et al., 2018, Mahato et al., 2018).

4. Differential sensitivity in inverse problems, sparse representations, and radar image formation

In quantitative phase imaging and dielectric parameter estimation, IS is formalized as a field derivative. If the scattered field is $2$00 and detector measurements are $2$01, then the Sensitivity Field is

$2$02

It enters the gradient of the quadratic loss through

$2$03

Because $2$04 satisfies a Lippmann–Schwinger equation, $2$05 obeys the analogous relation

$2$06

with $2$07. The first term scatters the sensitivity field through the same potential, and the second acts as a source induced by $2$08 (Barbastathis, 2024).

In sparse coding, IS is the gain from an image perturbation $2$09 to the induced code perturbation $2$10:

$2$11

If $2$12, then for a unit perturbation $2$13,

$2$14

and the maximal sensitivity is the largest singular value of $2$15. For the $2$16 sparse code defined by

$2$17

the locally active coefficients satisfy

$2$18

The analysis attributes high sensitivity to linear combinations of active dictionary elements with high cancellation; empirically, on MNIST, sparse codes are most sensitive to elastic distortions, less to swapping, and least to noise, which is the opposite of the desideratum for invariant recognition (Luther et al., 2022).

In BPA SAR image formation, sensitivity is derived with respect to initial position, velocity, and attitude navigation errors. The study develops Taylor-expanded range expressions showing that position errors mainly shift the BPA curve, while along-track velocity errors and pitch or roll attitude errors induce quadratic phase mismatch and hence azimuth blur; yaw errors do not affect a straight-and-level BPA image to first order. Simulation and real-data experiments show that the predicted shift and blur agree with observed image changes to within one pixel (Lindstrom et al., 2020).

Taken together, these formulations treat image sensitivity as a derivative-mediated propagation law. In one case the derivative is with respect to physical object parameters, in another with respect to input images or latent coefficients, and in another with respect to navigation state. The formal object varies, but the organizing idea is a local linearization of a nonlinear image formation or encoding map.

5. Counterfactual and variance-based sensitivity of classifiers

For facial analysis and related vision systems, image counterfactual sensitivity analysis measures how a classifier changes under controlled semantic edits produced by a generative model. Given a generator $2$19, classifier $2$20, binary decision $2$21, and unit-norm attribute vector $2$22, score sensitivity is

$2$23

and classification sensitivity is

$2$24

The proof-of-concept study trains a Progressive GAN on $2$25 CelebA images, infers attribute vectors by fitting linear SVMs or logistic regressions in latent space, and audits a smiling classifier with overall accuracy $2$26. For $2$27 versus $2$28, Heavy_Makeup produces a $2$29 score shift and a $2$30 flip rate of about $2$31, No_Beard raises smiling score by about $2$32, Young affects predictions as well, and sensitivities nearly double near the decision boundary $2$33 (Denton et al., 2019).

A different black-box formulation models image perturbations as random variables and quantifies their effect through Sobol indices computed via generalized polynomial chaos. With a fixed clean image $2$34, perturbation variables $2$35, and a transform $2$36, the classifier output is mapped to

$2$37

A polynomial chaos expansion

$2$38

then yields a variance decomposition and first-order or higher-order Sobol indices. In a welding-defect case study using a fine-tuned ResNet18, the relative Sobol indices are reported as $2$39, $2$40, $2$41, and joint interactions summing to about $2$42. In a BMW emblem classification problem, brightness yields $2$43, rotation yields $2$44, and the brightness–rotation interaction $2$45 is dominant (Bahr et al., 23 Jun 2025).

These approaches operationalize IS for predictive systems that may be treated as black boxes. One emphasizes controlled semantic counterfactuals in latent space; the other emphasizes distributional domain shifts and variance attribution. In both cases, large sensitivity identifies perturbations that most compromise model reliability.

6. Perceptual and similarity-based sensitivity

In perceptual image quality analysis, sensitivity can be defined directly from a full-reference distortion metric. For a reference patch $2$46 and distorted patch $2$47 with $2$48, the sensitivity of a perceptual metric $2$49 is

$2$50

Using PixelCNN++ to estimate $2$51, the study evaluates probability-related factors and finds that two are especially informative: $2$52 and the RMS contrast $2$53. A two-factor regression model,

$2$54

achieves correlation $2$55 with subjective metrics in the abstract and an average Pearson correlation $2$56 across MS-SSIM, NLPD, PIM, LPIPS, and DISTS in the detailed summary; a Random-Forest regressor on the same factors reaches $2$57. The model is further validated by reproducing the Contrast Sensitivity Function, Weber’s law, and contrast masking trends (Hepburn et al., 2023).

A separate line of work concerns the sensitivity of similarity indexes themselves for low-information scientific images. Starting from standard SSIM,

$2$58

two alternatives are proposed: Intensity-Weighted SSIM (ITW-SSIM), which replaces SSIM moments by weighted versions built from a monotone intensity weighting function $2$59, and the Low-Information Similarity Index (LISI), which emphasizes pixels that are both high-intensity and nearly equal. LISI is defined as

$2$60

with $2$61 and $2$62. The auxiliary sensitivity index

$2$63

measures how much faster a proposed metric drops than baseline SSIM, and the direction index is $2$64. On radio-astronomical images with $2$65 versus $2$66 noise, one example gives $2$67, $2$68, $2$69, $2$70, and $2$71, with corresponding sensi values of about $2$72, $2$73, $2$74, and $2$75 (Li et al., 2022).

A plausible implication is that perceptual and similarity-based notions of IS occupy a different level of abstraction from detector or forward-model sensitivity. They do not ask how much signal a sensor records, but how strongly a comparison rule or perceptual surrogate responds to distortions of fixed physical magnitude. Even so, they preserve the same underlying logic: response magnitude is normalized by a specified perturbation model, whether that perturbation is additive noise, semantic latent motion, brightness–rotation drift, or sub-pixel displacement.