Infrared Super-Resolution GAN (IR-SRGAN)

• IR-SRGAN is a family of GAN-based super-resolution methods that enhance low-resolution infrared images by incorporating domain-specific noise, contrast, and thermal fidelity constraints.
• The framework customizes generator and discriminator architectures with specialized loss functions and residual designs to preserve critical spectral and structural details.
• IR-SRGAN integrates visible-domain priors with sensor-specific adaptations, leading to improvements in quality metrics like PSNR and SSIM and boosting downstream tasks such as detection and segmentation.

Searching arXiv for the cited infrared super-resolution papers to ground the article in current sources. arXiv search query: "all:Infrared Image Super-Resolution via Heterogeneous Convolutional WGAN" arXiv search query: "all:Texture and Noise Dual Adaptation for Infrared Image Super-Resolution" arXiv search query: "all:Contourlet Refinement Gate Framework for Thermal Spectrum Distribution Regularized Infrared Image Super-Resolution" Infrared Super-Resolution Generative Adversarial Network (IR-SRGAN) denotes a family of GAN-based super-resolution models that adapt SRGAN/ESRGAN-style architectures to infrared and thermal imagery, rather than a single canonical network (Huang et al., 2022). In this literature, the central task is single-image infrared super-resolution: given a low-resolution infrared image, reconstruct a high-resolution infrared image that is sharper and more useful for human or machine interpretation, while remaining compatible with infrared-specific image statistics, sensor noise characteristics, and, in some settings, thermal or spectral fidelity (Huang et al., 2021).

1. Scope, formulation, and infrared-specific constraints

The generic super-resolution problem is commonly written as

$I_{LR}=\mathbb{D}(I_{HR};\delta),$

with degradation operator

$$\mathbb{D}\left(I_{HR} ; \delta\right)=\left(I_{HR} \otimes \kappa\right) \downarrow_d+n_{\varsigma},$$

where $\kappa$ denotes blur, $\downarrow_d$ denotes downsampling, and $n_{\varsigma}$ denotes additive noise (Huang et al., 2022). Within IR-SRGAN, this formulation is retained, but the degradation is not treated as a simple RGB analogue. Infrared images are usually single-channel, exhibit low contrast and fewer high-frequency textures, and are affected by fixed-pattern noise, nonuniformity, temporal drift, limited optics, and coarse detector pitches (Huang et al., 2022).

The infrared setting introduces additional physical asymmetries. IR sensors capture thermal radiation rather than reflected visible light; IR images have longer wavelengths, fewer high spatial frequency components, and are more sensitive to the presence or absence of high-frequency information than RGB (Zou et al., 2024). As a consequence, direct transfer of RGB-oriented super-resolution models can distort the infrared spectral distribution, change the balance between high and low frequencies, hallucinate details, or oversmooth thermal boundaries, thereby degrading downstream infrared detection, segmentation, or tracking (Zou et al., 2024).

These constraints explain why IR-SRGAN research repeatedly departs from visible-image SR in three ways. First, generators are usually redesigned or regularized for weak-texture imagery. Second, adversarial training is often stabilized or weakened to limit hallucination. Third, evaluation is frequently extended beyond PSNR to SSIM, no-reference quality measures, or downstream task metrics such as detection and segmentation (Huang et al., 2022).

2. Architectural patterns

Most IR-SRGAN systems preserve the generator–discriminator decomposition of SRGAN, but specialize both components for thermal imagery, domain adaptation, or application-specific structure. Representative variants are summarized below.

System	Adversarial setup	Distinguishing mechanism
HetSRWGAN (Huang et al., 2021)	WGAN	Heterogeneous kernel-based residual blocks and gradient cosine similarity loss
DASRGAN (Huang et al., 2023)	Standard BCE-GAN	Texture-Oriented Adaptation and Noise-Oriented Adaptation
Transfer-learning IR-SRGAN for FM-PCT (Zhu et al., 13 Sep 2025)	GAN with U-Net discriminator	ESRGAN-like generator adapted to thermographic tomograms
Cross-modal pathology GAN (Falahkheirkhah et al., 2019)	Conditional GAN	H&E-supervised morphology transfer with conditional normalization

A recurrent generator design is the residual SR backbone. In HetSRWGAN, the core block is the heterogeneous kernel-based residual block, which replaces a homogeneous $3\times 3$ convolution with heterogeneous convolution and retains residual propagation: $$\mathbf{x}_{i+1} = \mathcal{F}(\mathbf{x}_i, \mathcal{W}_i) + \mathbf{x}_i.$$ This design is explicitly lightweight: HetSRWGAN reports $0.496$M parameters and $0.095$G FLOPs, versus $0.956$M parameters and $$\mathbb{D}\left(I_{HR} ; \delta\right)=\left(I_{HR} \otimes \kappa\right) \downarrow_d+n_{\varsigma},$$0G FLOPs for the referenced SRGAN baseline configuration (Huang et al., 2021). Batch normalization is removed in both generator and discriminator, following the observation that BN can hurt super-resolution and destabilize adversarial learning in IR settings (Huang et al., 2021).

DASRGAN instead adopts an ESRGAN / Real-ESRGAN style RRDB generator and concentrates most infrared specialization in the discriminator and losses (Huang et al., 2023). Its stage-two discriminator contains a main branch and a Sobel-based prior extraction branch,

$$\mathbb{D}\left(I_{HR} ; \delta\right)=\left(I_{HR} \otimes \kappa\right) \downarrow_d+n_{\varsigma},$$1

so that texture priors extracted from gradients participate directly in adversarial discrimination (Huang et al., 2023).

Application-specific IR-SRGANs often depart further from the baseline SRGAN topology. In photothermal coherence tomography for impact-damage inspection, the generator follows an ESRGAN-like design with 23 Residual-in-Residual Dense Blocks, while the discriminator is U-Net-based and combines a global encoder decision with pixel-wise decoder supervision (Zhu et al., 13 Sep 2025). In infrared spectroscopic microscopy, a conditional GAN uses a U-Net + ResNet generator with three down-sampling blocks, one bridge block, three up-sampling blocks, and conditional normalization driven by tissue-class masks, producing super-resolved morphology from low-resolution single-band IR input (Falahkheirkhah et al., 2019).

3. Objective functions and regularization

IR-SRGAN loss design is typically more elaborate than in RGB SRGAN because edge sharpness, noise suppression, thermal plausibility, and distribution matching are not aligned objectives. The survey literature therefore places GAN-based IR SR alongside pixel losses, perceptual losses, SSIM-type structural terms, and explicit edge regularizers (Huang et al., 2022).

HetSRWGAN combines content loss, Wasserstein adversarial loss, and a gradient cosine similarity term: $$\mathbb{D}\left(I_{HR} ; \delta\right)=\left(I_{HR} \otimes \kappa\right) \downarrow_d+n_{\varsigma},$$2 with $$\mathbb{D}\left(I_{HR} ; \delta\right)=\left(I_{HR} \otimes \kappa\right) \downarrow_d+n_{\varsigma},$$3 and $$\mathbb{D}\left(I_{HR} ; \delta\right)=\left(I_{HR} \otimes \kappa\right) \downarrow_d+n_{\varsigma},$$4 (Huang et al., 2021). The cosine similarity is computed on flattened image-gradient representations,

$$\mathbb{D}\left(I_{HR} ; \delta\right)=\left(I_{HR} \otimes \kappa\right) \downarrow_d+n_{\varsigma},$$5

so minimizing $$\mathbb{D}\left(I_{HR} ; \delta\right)=\left(I_{HR} \otimes \kappa\right) \downarrow_d+n_{\varsigma},$$6 aligns the orientation and relative magnitude of reconstructed and target edges (Huang et al., 2021). This is particularly relevant in IR, where boundaries often carry more semantic weight than texture fields.

DASRGAN separates adaptation into texture and noise channels. Its texture-oriented prior loss is an $$\mathbb{D}\left(I_{HR} ; \delta\right)=\left(I_{HR} \otimes \kappa\right) \downarrow_d+n_{\varsigma},$$7 penalty on Sobel magnitudes,

$$\mathbb{D}\left(I_{HR} ; \delta\right)=\left(I_{HR} \otimes \kappa\right) \downarrow_d+n_{\varsigma},$$8

with

$$\mathbb{D}\left(I_{HR} ; \delta\right)=\left(I_{HR} \otimes \kappa\right) \downarrow_d+n_{\varsigma},$$9

while its noise-oriented loss pushes the generated result away from synthetic degradation patterns in VGG feature space,

$\kappa$0

The aggregated optimization is written as

$\kappa$1

with the best reported setting $\kappa$2 (Huang et al., 2023).

A major counterpoint in this literature is that adversarial realism is not identical to infrared fidelity. The contourlet-based CoRPLE framework is explicitly non-GAN and introduces a Spectral Fidelity Loss,

$\kappa$3

where the normalized terms are obtained from the log-magnitude 2D DFT spectra of HR and SR images after DC centering (Zou et al., 2024). The same work states that, for IR-SRGAN practitioners, $\kappa$4 can be added to a conventional GAN objective,

$\kappa$5

to discourage the generator from altering infrared frequency statistics (Zou et al., 2024). This suggests that the mature form of IR-SRGAN is not purely adversarial; it is adversarially trained but physically constrained.

4. Domain adaptation, conditioning, and cross-modal supervision

A prominent strand of IR-SRGAN research addresses the scarcity of high-quality infrared training data by importing priors from visible images. The key difficulty is that visible imagery provides richer texture but also carries incompatible noise and blur statistics. DASRGAN formulates this explicitly through triplets $\kappa$6, where $\kappa$7 is the visible-RGB LR counterpart and all LR images are obtained via bicubic downsampling (Huang et al., 2023). Its training is two-stage: IR-only SR pre-training with MAE, followed by target-oriented domain adaptation in which the pre-trained generator processes visible LR input and is forced toward the IR HR distribution under the control of Texture-Oriented Adaptation and Noise-Oriented Adaptation (Huang et al., 2023).

The same transfer-learning logic appears in other forms across the literature. The survey describes progressive SR GANs in which a main path is trained on large visible datasets and an IR branch is adapted afterward, as well as multimodal systems that fuse RGB and thermal cues to restore detail that is absent in the native infrared channel (Huang et al., 2022). Unpaired or weakly paired thermal SR is also represented through CycleGAN-like domain-transfer designs with cycle consistency and Sobel-based contour preservation (Huang et al., 2022).

Conditional supervision can be stronger than simple visible guidance. In the pathology-oriented GAN, the generator is conditioned by spatial class masks for stroma, epithelium, and others through conditional normalization,

$\kappa$8

so that reconstructed high-resolution morphology is tied to tissue class rather than only to low-resolution intensity patterns (Falahkheirkhah et al., 2019). Although this system is trained with H&E as the HR teacher rather than with HR IR images, it is still structurally an infrared super-resolution GAN in the sense that real low-resolution IR input is mapped to a super-resolved output by adversarial and perceptual training (Falahkheirkhah et al., 2019).

These developments indicate that “IR-SRGAN” increasingly denotes a conditional or adapted adversarial framework, not merely a visible-image SRGAN applied to thermal data.

5. Evaluation protocols and application domains

Evaluation practice in IR-SRGAN spans classical image-reconstruction scores and task-oriented criteria. Standard metrics include PSNR, SSIM, MSE, NIQE, LPIPS, and, in some survey discussions, composite no-reference quality measures such as $\kappa$9 (Huang et al., 2022). Because infrared SR often supports machine perception, several works also evaluate object detection or semantic segmentation performance after super-resolution (Zou et al., 2024).

On fused infrared datasets, HetSRWGAN reports PSNR $\downarrow_d$0 and SSIM $\downarrow_d$1 on fusionA-22, and PSNR $\downarrow_d$2 and SSIM $\downarrow_d$3 on fusionC-22, while maintaining the lightweight complexity noted earlier (Huang et al., 2021). The same study emphasizes that SRGAN on its IR setup can suffer mode collapse, whereas the Wasserstein version is more stable (Huang et al., 2021).

On paired infrared–visible street-scene data, DASRGAN reports state-of-the-art performance on most tested settings. For example, on M3FD15 at $\downarrow_d$4, it reports PSNR $\downarrow_d$5, MSE $\downarrow_d$6, and SSIM $\downarrow_d$7; on M3FD15 at $\downarrow_d$8, it reports PSNR $\downarrow_d$9, MSE $n_{\varsigma}$0, and SSIM $n_{\varsigma}$1 (Huang et al., 2023). Cross-dataset generalization is also reported on CVC, with DASRGAN outperforming comparison models on both $n_{\varsigma}$2 and $n_{\varsigma}$3 settings (Huang et al., 2023). Its ablations show that naive visible-domain adaptation can degrade performance, while target-oriented adaptation yields substantial gains (Huang et al., 2023).

The non-GAN CoRPLE framework is relevant because it demonstrates that infrared SR should be judged beyond visual plausibility. Using SR outputs as input to YOLOv5 and SegFormer, it reports $n_{\varsigma}$4 detection mAP $n_{\varsigma}$5 versus a best competitor near $n_{\varsigma}$6, $n_{\varsigma}$7 detection mAP $n_{\varsigma}$8 versus a best competitor near $n_{\varsigma}$9, $3\times 3$0 semantic-segmentation mIoU $3\times 3$1 versus a best competitor near $3\times 3$2, and $3\times 3$3 mIoU $3\times 3$4 versus a best competitor $3\times 3$5 (Zou et al., 2024). A plausible implication is that IR-SRGAN evaluation is incomplete if it ignores whether the reconstructed thermal structure remains useful for downstream inference.

Application-specific IR-SRGANs extend the domain far beyond surveillance. In nondestructive evaluation of impact damage in CFRP and CFRTP laminates, a transfer-learning IR-SRGAN is applied slice-wise to FM-PCT thermograms to sharpen crack and delamination boundaries; the study emphasizes qualitative gains in lateral clarity, but does not report PSNR, SSIM, LPIPS, or other numerical SR metrics (Zhu et al., 13 Sep 2025). In IR spectroscopic imaging for pathology, conditional GAN super-resolution is used to recover morphology beyond the diffraction limit while retaining IR-like contrast at a single wavenumber band, with the output intended to approach H&E-scale structural detail (Falahkheirkhah et al., 2019).

6. Limitations, misconceptions, and research directions

Several misconceptions recur in discussions of IR-SRGAN. The first is terminological: IR-SRGAN is not a standardized single model, but a moving class of GAN-based infrared SR systems (Huang et al., 2022). The second is bibliographic: a chapter titled “Infrared Image Super-Resolution via GAN” does not, in the cited source, contain a technical model description, loss design, or experiments, and therefore does not define a canonical IR-SRGAN architecture (Huang et al., 2023). The third is methodological: adversarial sharpening alone is not sufficient for infrared SR, because visually sharp outputs can still distort thermal statistics, spectral distribution, or downstream detector performance (Zou et al., 2024).

The limitations reported across the literature are consistent. Real infrared degradations are more complex than bicubic downsampling; contourlet or wavelet decompositions increase complexity and computational cost; RGB-trained perceptual or LLMs may not fully align with thermal imagery; paired visible–infrared data are often required for adaptation; and some application papers remain primarily qualitative (Zou et al., 2024). GAN variants add their own liabilities, including mode collapse, artifact generation, and the risk of introducing non-physical high-frequency content (Huang et al., 2021).

The most credible research directions therefore combine, rather than replace, adversarial training with stronger domain structure. The survey emphasizes physics-informed or sensor-integrated models, multimodal and cross-spectral SR, self-supervised or unsupervised adaptation, lightweight real-time networks, and newer generative paradigms such as transformers and diffusion models (Huang et al., 2022). The contourlet-based spectral-regularization line suggests that future IR-SRGAN systems are likely to remain adversarial only in part: generator realism will be anchored by explicit frequency-domain, gradient-domain, semantic, or task-level constraints, so that super-resolved infrared imagery remains thermally meaningful as well as perceptually sharp (Zou et al., 2024).