InverseNet: Benchmarking Operator Mismatch and Calibration Across Compressive Imaging Modalities

Chengshuai Yang; Xin Yuan

cs.CVMar 4, 2026

InverseNet: Benchmarking Operator Mismatch and Calibration Across Compressive Imaging Modalities

Authors: Chengshuai Yang, Xin Yuan

Organizations: NextGen PlatformAI C Corp, USA · School of Engineering, Westlake University, Hangzhou, China

Abstract

State-of-the-art EfficientSCI loses 20.58 dB when its assumed forward operator deviates from physical reality in just eight parameters, yet no existing benchmark quantifies operator mismatch, the default condition in deployed compressive imaging systems. We introduce InverseNet, the first cross-modality benchmark for operator mismatch, spanning CASSI, CACTI, and single-pixel cameras. Evaluating 12 methods under a four-scenario protocol (ideal, mismatched, oracle-corrected, blind calibration) across 27 simulated scenes and 9 real hardware captures, we find: (1) deep learning methods lose 10-21 dB under mismatch, eliminating their advantage over classical baselines; (2) performance and robustness are inversely correlated across modalities (Spearman r_s = -0.71, p < 0.01); (3) mask-oblivious architectures recover 0% of mismatch losses regardless of calibration quality, while operator-conditioned methods recover 41-90%; (4) blind grid-search calibration recovers 85-100% of the oracle bound without ground truth. Real hardware experiments confirm that simulation trends transfer to physical data. Code will be released upon acceptance.

Figures & tables

Figure 1: PSNR across three scenarios for all modalities. Scenario I (Ideal): perfect operator. Scenario II (Baseline): mismatched operator. Scenario III (Oracle): true operator used for reconstruction. The collapse of deep learning methods under Scenario II is visible across all modalities, with CACTI showing the most severe degradation.

Figure 2: Qualitative reconstruction comparison across three modalities. Each row shows a representative scene for one modality (CASSI: Scene 1 band 14, MST-L; CACTI: kobe frame 4, ELP-Unfolding; SPC: cameraman , HATNet). Error maps (jet colormap, same scale per row) highlight how mismatch (Scenario II) introduces spatially structured artifacts that oracle correction (Scenario III) largely removes.

Method	Scenario I	Scenario II	Scenario III	$\boldsymbol{\Delta}_{\textbf{deg}}$	$\boldsymbol{\Delta}_{\textbf{rec}}$	$\boldsymbol{\rho}$
GAP-TV [ 7 ]	24.34 $\pm$ 1.90 / .722	20.96 $\pm$ 1.62 / .611	21.72 $\pm$ 1.48 / .687	3.38	0.76	22.5%
PnP-HSICNN [ 33 ]	25.12 $\pm$ 1.88 / .758	20.40 $\pm$ 1.71 / .574	23.08 $\pm$ 1.52 / .702	4.72	2.68	56.8% †
HDNet [ 11 ]	34.66 $\pm$ 2.62 / .970	21.88 $\pm$ 1.72 / .756	21.88 $\pm$ 1.72 / .756	12.78	0.00	0%
MST-L [ 10 ]	34.81 $\pm$ 2.11 / .973	20.83 $\pm$ 2.01 / .744	27.33 $\pm$ 1.86 / .881	13.98	6.50	46.5%

Table 1: CASSI reconstruction results (10 KAIST scenes,

256\times 256\times 28

, 5-parameter mismatch). PSNR (dB) / SSIM reported as mean

\pm

std.

\Delta_{\text{deg}}

: degradation (I

\rightarrow

II).

\Delta_{\text{rec}}

: oracle recovery (II

\rightarrow

III).

\rho

: recovery ratio. Best Scenario I and III PSNR in bold ; best

\rho

marked with

\dagger

Figure 3: Mismatch degradation (

\Delta_{\text{deg}}

, left) and oracle recovery (

\Delta_{\text{rec}}

, right) per method across all three modalities. CACTI suffers the most severe degradation (up to 20.6 dB) but also the highest absolute recovery. For CASSI, HDNet shows zero recovery due to its mask-oblivious architecture; PnP-HSICNN achieves the best recovery ratio (

\rho=56.8\%

). For SPC, HATNet recovers 10.4 dB (

\rho=89.6\%

Method	Scenario I	Scenario II	Scenario III	$\boldsymbol{\Delta}_{\textbf{deg}}$	$\boldsymbol{\Delta}_{\textbf{rec}}$	$\boldsymbol{\rho}$
GAP-TV [ 7 ]	26.75 $\pm$ 4.48 / .848	15.81 $\pm$ 1.98 / .305	26.01 $\pm$ 3.72 / .794	10.94	10.21	93.3%
PnP-FFDNet [ 16 ]	29.28 $\pm$ 5.53 / .890	11.43 $\pm$ 2.71 / .216	25.39 $\pm$ 3.52 / .820	17.85	13.96	78.2%
ELP-Unfolding [ 13 ]	34.09 $\pm$ 4.11 / .965	15.47 $\pm$ 1.71 / .308	29.40 $\pm$ 3.15 / .927	18.63	13.93	74.8%
EfficientSCI [ 12 ]	35.39 $\pm$ 4.46 / .973	14.81 $\pm$ 2.19 / .303	27.38 $\pm$ 3.52 / .927	20.58	12.57	61.1%

Table 2: CACTI reconstruction results (6 videos,

256\times 256\times 8

). PSNR (dB) / SSIM reported as mean

\pm

std. CACTI exhibits the most severe mismatch degradation of all three modalities.

Method	Scenario I	Scenario II	Scenario III	$\boldsymbol{\Delta}_{\textbf{deg}}$	$\boldsymbol{\Delta}_{\textbf{rec}}$	$\boldsymbol{\rho}$
FISTA-TV [ 8 ]	28.06 $\pm$ 3.38 / .852	18.51 $\pm$ 0.69 / .586	26.21 $\pm$ 2.28 / .759	9.55	7.71	80.7%
PnP-DRUNet [ 34 ]	30.53 $\pm$ 3.36 / .899	16.29 $\pm$ 0.75 / .415	23.65 $\pm$ 1.46 / .666	14.24	7.37	51.7%
ISTA-Net [ 14 ]	31.85 $\pm$ 3.11 / .916	19.02 $\pm$ 0.61 / .584	27.45 $\pm$ 1.32 / .760	12.83	8.43	65.7%
HATNet [ 15 ]	30.98 $\pm$ 0.95 / .847	19.40 $\pm$ 0.59 / .648	29.78 $\pm$ 0.81 / .807	11.58	10.38	89.6%

Table 3: SPC reconstruction results (11 Set11 images,

256\times 256

, 25% sampling). PSNR (dB) / SSIM reported as mean

\pm

std.

Modality	Dim.	$\boldsymbol{\Delta}_{\textbf{deg}}$ range	$\boldsymbol{\Delta}_{\textbf{rec}}$ range	$\boldsymbol{\rho}_{\textbf{best}}$	SSIM rec.	Best method
CASSI	5	.11–.23	.04–.21	69.6%	.574 $\to$ .702	PnP-HSICNN
CACTI	8	.54–.67	.04–.07	94.2%	.308 $\to$ .927	ELP-Unf.
SPC	2	.20–.48	.04–.23	80.1%	.648 $\to$ .807	HATNet

Table 4: Cross-modality summary using SSIM-based metrics (note: Tables 1 – 3 report PSNR-based values; SSIM is used here for cross-modality comparability as it normalises for different signal ranges). Dim. = number of mismatch parameters.

\Delta_{\text{deg}}

and

\rho

are computed from SSIM. SSIM rec. = best-method absolute SSIM change from Scenario II to III. PSNR-based

\rho

values for the best method per modality are: CASSI 56.8%, CACTI 93.3%, SPC 89.6%.

Figure 4: Recovery ratio (

\rho

) vs. ideal PSNR (Scenario I) for all 12 methods across three modalities. Color indicates modality; shape indicates method type (classical, operator-aware, mask-oblivious). An inverse trend is visible: higher-performing methods tend to have lower recovery ratios, suggesting that stronger learned priors create greater operator dependence.

Method	Calibrated	Mismatched	Ratio
GAP-TV	0.0019	0.0033	1.8 $\times$
PnP-HSICNN	0.0127	0.0142	1.1 $\times$

Table 5: CASSI real data (5 scenes,

660\times 660\times 28

). Normalised measurement residual

r=\|\mathbf{y}-\mathbf{\Phi}_{\text{cal}}\hat{\mathbf{x}}\|^{2}/\|\mathbf{y}\|^{2}

(lower is better; no ground truth required). Ratio: mismatched / calibrated. Only classical/PnP methods evaluated (no ground truth for deep learning supervision).

Figure 5: CASSI real data (Scene 1, band 14): calibrated (top) vs. mismatched (bottom) reconstructions for GAP-TV and PnP-HSICNN. The visual differences are subtle, consistent with the modest residual ratios in table 5 —spatial mask shift alone is not the dominant degradation source for CASSI.

Method	Calibrated	Mismatched	Ratio
GAP-TV	1.6e-5	1.6e-4	10.4 $\times$
PnP-FFDNet	0.0041	0.0069	2.0 $\times$

Table 6: CACTI real data (4 scenes,

512\times 512

, cr

=

10). Normalised measurement residual (lower is better). Ratio: mismatched / calibrated.

Figure 6: CACTI real data: calibrated vs. mismatched reconstruction for duomino and hand scenes. Mismatched masks produce temporal ghosting artifacts, mirroring simulation findings.

Method	II	IV	III	Recovery
FISTA-TV	19.78	26.54	27.60	86%
PnP-DRUNet	18.34	25.39	26.01	92%

Table 7: SPC blind calibration via TV minimisation ( eq. 9 ). II: mismatched, no calibration. IV: estimated

\hat{\alpha}

via TV grid search. III: oracle (true

\alpha

). Recovery: (IV

-

II) / (III

-

II). True

\alpha{=}0.0015

; estimated

\hat{\alpha}{=}0.00125

Computational Imaging

Computational imaging combines optical and sensor design with algorithms to recover or infer scene information from indirect, compressed, or noisy measurements.

See all 123 papers

A neural operator view on U-Nets for inverse imaging problems
Aug 6, 2026Alexander Auras, Martin Burger, Samira Kabri +2U-Net Neural Operators
Computer Vision Group, University of Siegen, Hölderlinstraße 3, 57076 Siegen, Germany · Helmholtz Imaging, Deutsches Elektronen-Synchrotron DESY, Notkestr. 85, Hamburg, 22607, Germany · Fachbereich Mathematik, Universität Hamburg, Bundesstrasse 55, Hamburg, 20146, Germany
The Finite Primitive Basis Theorem for Computational Imaging: Formal Foundations of the OperatorGraph Representation
Feb 24, 2026Chengshuai YangComputational Imaging World Models
NextGen PlatformAI C Corp, USA
Resolving Blind Inverse Problems under Dynamic Range Compression via Structured Forward Operator Modeling
Mar 2, 2026Muyu Liu, Xuanyu Tian, Chenhe Du +3Image Inverse Problems HDR Imaging
ShanghaiTech University · Shanghai Jiao Tong University

Aug 6, 2026J/K move · Enter open · S save

A neural operator view on U-Nets for inverse imaging problems

Alexander Auras, Martin Burger, Samira Kabri, Michael Moeller+1

Computer Vision Group, University of Siegen, Hölderlinstraße 3, 57076 Siegen, Germany · Helmholtz Imaging, Deutsches Elektronen-Synchrotron DESY, Notkestr. 85, Hamburg, 22607, Germany · Fachbereich Mathematik, Universität Hamburg, Bundesstrasse 55, Hamburg, 20146, Germany

U-Net Neural Operators Image Inverse Problems

Open paper AskPDF arXiv

Deep neural networks have shown great empirical success in the solution of a wide variety of ill-posed inverse problems in imaging. Yet, very few works have studied their behavior in the limit that turns the discretized ill-conditioned problems into truly ill-posed ones, i.e., for an increasing resolution of the discretization. In this work, we review common approaches to neural operator learning in architectures that resemble a U-Net, one of the most common classical architectures for inverse imaging problems. We discuss advantages and drawbacks of the respective approaches, consider a 1D toy example for improved interpretability, and present extensive numerical experiments on how different types of neural operator U-Nets can improve a first (crude) limited angle CT-reconstruction. In particular, we study how well networks trained for a certain resolution of the discretization generalize to other resolutions. Our finding is that while U-shaped neural operator architectures are by design resolution-invariant, the classical U-Net architecture seems to be more robust with respect to resolution changes than expected.

Abstract

Computational Imaging

Explore similar work

A neural operator view on U-Nets for inverse imaging problems

The Finite Primitive Basis Theorem for Computational Imaging: Formal Foundations of the OperatorGraph Representation

Resolving Blind Inverse Problems under Dynamic Range Compression via Structured Forward Operator Modeling