acceptodds
Withdrawn from ICLR 2027

Toward Physically Grounded Noise Modeling and Restoration for Uncooled Thermal Imaging

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Uncooled microbolometer cameras provide compact and low-cost thermal imaging, but their measurements are dominated by complex noise fluctuations. Accurately decomposing these physical noise sources is inherently ill-posed, as many components are strongly correlated and cannot be uniquely identified from limited calibration observations. We propose an identifiability-aware framework for noise modeling and restoration in uncooled thermal imaging. Instead of estimating all physical noise terms, we identify a minimal set of calibration parameters from blackbody flat-field sequences, including per-pixel gain and offset, while temperature/state variation and unresolved correlated components are represented through calibrated perturbation and residual sampling. The remaining correlated components are modeled by residual sampling, enabling physics-consistent noise synthesis. This synthesis process allows restoration networks to be trained using paired data generated from calibrated sensor measurements. Experiments on two VOx microbolometer cameras demonstrate improved restoration quality across distinct camera operating regimes compared with traditional and learning-based restoration baselines. Additional validation using high-frame-rate sequences captured by a cooled LWIR camera and SAM-based segmentation further supports the realism of the synthesized noise and the effectiveness of the proposed pipeline.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
thermal imaging, image restoration, noise synthesis, microbolometer sensors, fixed-pattern noise, physics-informed modeling

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_2j6JOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a calibration-driven noise synthesis and single-frame restoration method for uncooled thermal imaging. The authors estimate effective gain and offset through multi-temperature blackbody calibration, generate training pairs using measured residual sampling and random parameter perturbations, and train a lightweight U-Net for image restoration. Experiments on two uncooled thermal cameras demonstrate improved restoration performance, while reference data from a cooled camera and segmentation results provide additional validation.

Critical strengths
  1. Using measured sensor residuals to preserve structured noise is a practical and valuable approach to generating training data.

  2. The authors conduct multi-temperature calibration and collect data under two camera operating modes. The fine-tuned models achieve clear improvements in restoration metrics, as shown in the main results table.

Critical weaknesses
  1. Missing comparisons with relevant noise models. The paper does not compare its generator with structured infrared noise models from Cai et al. (TIP 2024) or PPFN (NeurIPS 2025). The Gaussian-noise ablation in Table 3 is insufficient to establish an advantage over these models under matched training conditions.

  2. Missing recent thermal restoration baselines. Tables 1 and 2 omit relevant methods, including PPFN, DEAL (CVPR 2025), ECMRNet (ICML 2026), and SEGD (CVPR Findings 2026). Comparisons under matched restoration settings would help establish the method’s advantages over these approaches.

  3. Insufficient support for identifiability. Section 3.1 does not specify conditions that separate gain from temperature-dependent drift, leaving the claimed identifiability of gain and offset unproven. Appendix B demonstrates the failure of an independence-based estimator under correlated noise, rather than a general non-identifiability result.

  4. Unclear Camera 2 data split. Section 4.1 reports 11 training sequences and five validation sequences, whereas Table 2 evaluates 11 scenes without explaining their relationship. This ambiguity prevents verification that the evaluation scenes are independent of the training data.

  5. Unclear attribution of fine-tuning gains. Sections 4.3, 4.6, and Appendix D do not clearly distinguish synthetic pairs from real supervision during fine-tuning, which yields gains of 2.65 and 3.45 dB. Table 3 also leaves unclear whether all ablations use identical fine-tuning data and budgets, making the contribution of noise modeling difficult to isolate.

  6. No independent validation of noise realism. Appendix E evaluates synthetic data produced by the same generation process used for training, which cannot independently establish realism. Quantitative comparisons with measured noise statistics under held-out temperatures or operating conditions are missing.

  7. Formatting and presentation issues. The title and section headings do not follow the small-caps style specified by the official ICLR 2027 template. Tables 1 and 2 also fail to highlight the best values in bold, despite the text stating otherwise.

Questions for authors

The same as the weaknesses.

Reviewer_LN4MOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper revisits noise modeling for uncooled microbolometer thermal imaging from an identifiability perspective. The authors argue that a full physical decomposition of microbolometer noise is ill-posed under limited flat-field calibration, because the corresponding design matrix is rank-deficient and multiple parameter configurations explain the same observations. Instead, they adopt a minimal physics-informed formulation: per-pixel gain and offset are estimated from multi-temperature blackbody flat-field measurements, while the remaining structured residuals are treated as an empirical residual library sampled during synthesis.

Building on this, the paper proposes a calibration-and-sampling pipeline: calibrated non-uniformity correction (NUC) followed by residual noise sampling, with temperature-dependent offset perturbations, to synthesize realistic noisy-clean pairs via an inverse-forward NUC procedure. A lightweight U-Net is trained on these synthetic pairs with a two-stage protocol (pretraining on grayscale SID images with synthesized degradations, then camera-specific fine-tuning). Since uncooled sensors cannot provide high-quality supervision, the authors additionally collect high-frame-rate (1000 fps) cooled LWIR sequences as references.

Critical strengths
  • The paper does not simply propose another denoising network, instead, physically grounded noise modeling is meaningful for thermal sensor denoising.
  • The calibration-and-sampling pipeline is practical and the cooled-reference benchmark is a meaningful resource.
Critical weaknesses

(1) Method

  • The methodological contribution is an engineering combination of mature paradigms, Calibration-based noise synthesis + residual sampling + lightweight supervised restoration is a well-established recipe in CMOS RAW denoising. For ICLR, this raises the question of whether the contribution is a thermal-imaging application paper rather than a methodological advance.
  • There are inconsistencies in symbol usage around Eqs. (7)–(9) and the description of data sources during fine-tuning is unclear.

(2) Experiments

  • Experimental coverage and validation are insufficient to support the generalization and downstream claims. The evaluation uses only two VOx cameras with 15 and 11 scenes, which is a small test set for claims of consistent superiority across sensors.
  • The cooled LWIR reference differs from uncooled sensors in noise mechanism, spectral response, and spatial resolution, so its use as a clean reference is not fully justified.
  • The "clean reference" is constructed via temporal averaging + BM3D with a BRISQUE-selected noise level, but BM3D is also a compared baseline and is not a true ground truth; this creates a bias in both training supervision and evaluation.
  • Ablations cover only four variants on Camera 1, without sensitivity analysis for key hyperparameters.
  • Downstream SAM2 segmentation is purely qualitative, with no ground-truth masks or quantitative metrics, which is insufficient to support the conclusion that restoration improves high-level vision tasks.
Questions for authors
  • How to ensure the reliability of GT?
  • The evaluation covers only two VOx cameras (15 and 11 scenes), the generalization validation is Insufficient.
  • The difference between this paper and previous method for RGB raw sensor noise modeling.
Reviewer_9hbCOpenReview ↗
Rating
1
Confidence
1
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The submission deviates from the official ICLR 2027 template, with modifications to the font and the spacing between the title and abstract. These modifications may allow more content to fit within the page limit, creating an unfair advantage over other submissions. I am therefore inclined to reject this paper.

Critical strengths

N/A

Critical weaknesses

The submission deviates from the official ICLR 2027 template, with modifications to the font and the spacing between the title and abstract. These modifications may allow more content to fit within the page limit, creating an unfair advantage over other submissions. I am therefore inclined to reject this paper.

Questions for authors

N/A

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.