SFPRNet: A Spatio-Frequency Synergistic Progressive Restoration Network for Infrared Image Destriping

Highlights

What are the main findings?

• SFPRNet coordinates column-statistical stripe modeling, frequency-aware scale transition, and hierarchical skip refinement to suppress detector-induced stripe noise while preserving scene structures.
• The proposed method achieves competitive restoration performance on multi-source infrared data and consistently improves downstream infrared small-target detection performance.

What are the implications of the main findings?

• Task-specific spatial-frequency modeling provides a general design principle for balancing artifact suppression and structural preservation in infrared remote-sensing image restoration.
• Improving the quality of infrared remote-sensing imagery can benefit downstream AI-driven analysis, supporting destriping as an effective preprocessing step for reliable scene interpretation.

Abstract

Infrared stripe noise, mainly caused by detector nonuniformity and readout inconsistencies, is a common structured degradation in infrared imaging systems, exhibiting pronounced directional bias, strong column-wise persistence, and distinctive frequency-domain characteristics. These properties make generic restoration networks prone to a trade-off between stripe suppression and detail preservation: conventional two-dimensional attention often allocates modeling capacity to stripe-irrelevant spatial dependencies, early downsampling may entangle directional stripe components with scene structures, and skip connections in U-shaped architectures can reintroduce shallow residual stripe features into the decoder. Residual stripe artifacts and restoration-induced structural distortions can further impair downstream infrared image analysis, particularly small-target detection. To address these issues, we propose a Spatio-Frequency Synergistic Progressive Restoration Network (SFPRNet) for infrared image destriping. SFPRNet progressively exploits the column-wise statistical characteristics of stripe noise, directional frequency information during early scale transformation, and selective cross-level feature refinement to enhance stripe discrimination and suppression while preserving structural details. Extensive experiments on synthetic and real infrared images demonstrate that the proposed SFPRNet achieves superior or competitive performance across most datasets and degradation settings, providing a favorable balance between destriping quality and detail preservation. Furthermore, downstream evaluation with multiple infrared small-target detectors demonstrates that SFPRNet improves subsequent small-target detection performance.

1. Introduction

Infrared remote sensing is an important passive observation modality for spaceborne, airborne, and UAV-based platforms, as well as other long-range observation systems. Owing to its day-and-night imaging capability and reduced dependence on ambient illumination, it has been widely applied to environmental monitoring, disaster assessment, reconnaissance and early warning, and infrared small-target detection. In practical infrared remote-sensing systems, response inconsistency among detector elements, multi-channel readout mismatch, temperature drift, and residual radiometric-calibration errors can introduce fixed-pattern nonuniformity. A common manifestation is vertically distributed stripe noise, especially in infrared focal-plane-array and push-broom thermal imaging systems [1,2].

Unlike local random noise, infrared stripes exhibit pronounced directional bias, strong column-wise persistence, distinct structural correlation, and characteristic frequency-domain regularity. They not only reduce radiometric consistency and visual quality but may also distort scene structures and object boundaries, weaken thermal textures, and obscure dim target responses. Residual stripes and restoration-induced detail loss can further affect downstream remote-sensing tasks, including scene interpretation and infrared small-target detection [3,4,5]. Therefore, infrared destriping should suppress detector-induced nonuniformity while reliably preserving structural and target-related information.

Considerable effort has been devoted to infrared stripe suppression, ranging from calibration-based and scene-based nonuniformity correction to directional filtering, variational optimization, and deep learning-based restoration. Traditional techniques generally rely on explicit statistical assumptions, handcrafted directional constraints, or image priors. Although they can be effective under relatively regular degradation, their assumptions may become less reliable in complex remote-sensing scenes containing dense textures, weak targets, or nonuniform stripe distributions [1,2]. Deep models provide stronger nonlinear representation and reconstruction capabilities, but many of them still inherit architectural mechanisms from generic image denoising and restoration frameworks [3,4].

Such generic designs mainly emphasize general spatial correlation, larger receptive fields, or overall reconstruction fidelity, and are not fully aligned with the directional statistics, localized high-frequency behavior, and cross-level propagation characteristics of detector-induced infrared stripes. More recently, dedicated infrared destriping networks have incorporated directional dependencies, column-aware modeling, frequency-domain modeling, and specialized feature aggregation to better exploit the structured characteristics of stripe noise [3,4,6,7,8,9]. Despite these advances, several task-specific limitations remain in how current restoration architectures handle the structured characteristics of stripe noise. These limitations can be summarized as three challenges.

1.1. Limited Statistical Characterization of Stripe Responses

In generic restoration backbones, two-dimensional spatial attention or global feature interaction primarily models general spatial dependencies, whereas infrared vertical stripes are more strongly characterized by persistent column-wise response variations. Insufficient emphasis on these degradation-specific characteristics can make stripe patterns difficult to distinguish from genuine vertical structures. Recent stripe-aware methods have introduced directional and column-wise dependencies [3,4,9]; however, stripe characterization still relies largely on spatial feature relations, leaving the statistical variations induced by inter-column response inconsistency insufficiently characterized.

1.2. Loss of Directional Frequency Information During the First Scale Transition

Infrared stripe patterns are most explicit in shallow features, making the first scale transition particularly important for preserving their directional characteristics. Conventional strided convolution does not explicitly distinguish directional frequency components, and stripe-related responses may therefore become entangled with genuine structures during downsampling. Existing wavelet- and frequency-domain methods make directional sub-bands more explicit [9,10,11,12], but they mainly use these sub-bands for feature representation. By contrast, the specific role of the first scale transition and the adaptation of this transformation to the frequency composition of the input features remain comparatively underexplored.

1.3. Residual Stripe Leakage During Cross-Level Fusion

Skip connections in U-shaped architectures preserve high-resolution details, but they may also directly transmit residual stripe responses from the encoder to the decoder. Existing cross-level feature-selection and gated-fusion strategies mainly focus on generic feature filtering or detail recovery [8], whereas explicit suppression of stripe leakage through skip connections has received limited attention in infrared destriping. Moreover, the relative proportions of stripe residuals and useful structural information vary across feature scales, making identical treatment of different skip levels potentially suboptimal.

To address the above challenges, a Spatio-Frequency Synergistic Progressive Restoration Network (SFPRNet) is developed for infrared image destriping. SFPRNet follows a task-specific progressive restoration strategy tailored to the statistical, frequency, and cross-level characteristics of infrared stripe degradation. Specifically, Stripe-Aware Spatial Attention (SSA) establishes a column-statistical representation for discriminating stripe responses from genuine structures; High-Frequency Gated Sub-band Downsampling (HFGS-Down) formulates the first scale transition as an adaptive transformation conditioned on the frequency composition of shallow features; and Structure-Guided Skip Refinement (SSR) explicitly accounts for residual stripe transmission through skip connections and applies scale-dependent heterogeneous refinement. Together, these mechanisms strengthen stripe discrimination, preserve informative directional frequency characteristics during scale transformation, and suppress residual stripe reintroduction while maintaining structural fidelity.

Extensive experiments on multi-source synthetic datasets and self-acquired real infrared images demonstrate that SFPRNet achieves superior or competitive performance across different scenes and degradation settings, with a favorable balance between stripe suppression and structural detail preservation. Systematic ablation, robustness, and mechanism analyses further verify the effectiveness and stability of the proposed restoration strategy. In addition, comparative preprocessing experiments with multiple infrared destriping methods and representative small-target detectors demonstrate the practical value of the restored images for downstream analysis. Runtime, GPU-memory, and multi-resolution evaluations further show that SFPRNet maintains favorable computational efficiency and scalability for infrared image restoration.

The main contributions of this work are summarized as follows.

• A task-specific spatio-frequency progressive restoration framework, SFPRNet, is developed for infrared image destriping. The framework jointly considers the distinctive statistical characteristics, directional frequency behavior, and cross-level contamination of detector-induced stripes, providing a coordinated restoration strategy for stripe suppression and structural preservation.
• Three complementary mechanisms are introduced to realize this strategy. SSA performs degradation-specific column-statistical characterization to enhance stripe discrimination; HFGS-Down establishes a frequency-composition-aware adaptive transformation at the first scale transition; and SSR explicitly suppresses stripe leakage through scale-dependent heterogeneous refinement of cross-level features.
• Comprehensive experiments on multi-source synthetic and self-acquired real infrared data validate the restoration performance and generalization ability of SFPRNet. Systematic robustness and mechanism analyses, comparative downstream detection experiments, and computational-efficiency evaluations further demonstrate its stability and practical applicability.

2. Related Work

2.1. Traditional and Model-Based Infrared Destriping Methods

In infrared remote-sensing systems, stripe noise mainly originates from detector response mismatch, readout nonuniformity, and environment-induced system drift. Accordingly, early studies were largely developed from the perspective of nonuniformity correction (NUC). These methods can generally be divided into calibration-based approaches and scene-based approaches. The former estimate gain and bias parameters using blackbody references or multi-point calibration. They are easy to implement and computationally efficient, but they depend heavily on imaging conditions and hardware procedures, which limits their adaptability under varying remote-sensing observation conditions. The latter estimate nonuniformity parameters online by exploiting image statistics, frame redundancy, or scene motion, thereby avoiding extra calibration hardware. However, they usually rely on strong assumptions about motion or scene stationarity and may suffer from ghosting, stripe residuals, or estimation drift in static backgrounds or weak-texture regions [13,14].

To reduce the reliance on multi-frame motion assumptions, single-image infrared destriping has attracted increasing attention. Representative methods include adaptive grayscale adjustment, directional filtering, total-variation regularization, and image decomposition [15,16,17,18]. These methods explicitly encode stripe-direction priors or structural constraints to separate background content from stripe components, and they can work reasonably well under regular degradations and simple scenes. When real infrared images contain complex building textures, weak target edges, or dense high-frequency backgrounds, however, traditional methods often struggle to balance stripe suppression and structure preservation, and they tend to produce oversmoothing, blurred edges, or detail loss. More recent model-based methods have further considered adaptive regularization, local trend repair, and the joint separation of stripe and random noise to improve robustness in push-broom and other remote-sensing imagery [1,2,19].

2.2. Deep Learning-Based Infrared Destriping

With the rapid development of deep learning for image restoration, convolutional neural networks, residual learning, and U-shaped encoder-decoder architectures have been widely adopted for infrared destriping [12,20]. Compared with traditional methods, deep models can learn nonlinear mappings from degraded images to clean images in a data-driven manner, and therefore usually achieve stronger restoration performance on complex infrared scenes and large synthetic datasets. Recent studies have explored multi-scale modeling and attention mechanisms, leading to a series of dedicated deep destriping architectures. Representative designs have further incorporated column-spatial correction, stage-wise gain and offset estimation, adversarial learning, and asymmetric sampling [3,4,6,7]. These developments indicate that infrared destriping has gradually evolved from directly adopting generic restoration architectures to task-specific modeling of stripe degradation. Nevertheless, existing specialized designs still differ considerably in how explicitly the intrinsic characteristics of detector-induced stripes are represented and exploited during feature extraction, scale transformation, and cross-level feature fusion.

2.3. Stripe-Aware Modeling and Specialized Design

Given the strong directional bias of vertical stripes in infrared remote-sensing imagery, recent studies have introduced degradation-oriented modeling to improve stripe correction. CSCNet [3] exploits column-spatial correction to enhance directional consistency and image uniformity, while the Noise Parameter Estimation Two-Stage Network [4] models nonuniformity through degradation-parameter estimation and staged correction. More recently, ASCNet [7] employs a Column Nonuniformity Correction Module to capture global column relationships by jointly modeling column uniformity, spatial correlation, and self-dependence. Although these approaches substantially strengthen stripe-oriented modeling, their discrimination mechanisms are still primarily built upon correction relationships, estimated degradation parameters, or learned column-wise dependencies. The detector-induced statistical response characteristics of abnormal columns themselves are not explicitly formulated as complementary discriminative descriptors. Consequently, when stripe artifacts and genuine vertical structures exhibit similar spatial or contextual patterns, the available distinction still depends heavily on learned relational representations.

Frequency-domain and wavelet-based modeling provides another important route for stripe restoration. MWCNN [10] introduces multi-level wavelet decomposition into an encoder-decoder architecture for information-preserving resolution transformation, while SNRWDNN [12] learns stripe-noise representation and restoration in the wavelet domain. DestripeCycleGAN [6] further employs a Multi-level Wavelet U-Net in which Haar wavelets are used for sampling to reduce directional information loss. Despite these advances, wavelet decomposition in existing designs mainly functions as a predefined representation, restoration domain, or sampling operator. The frequency composition of intermediate features does not directly determine how the subsequent scale transformation should be performed. More importantly, multi-level wavelet processing is generally organized according to a predefined network hierarchy, without explicitly accounting for how the discriminative characteristics of stripe-related frequency responses evolve across different scale transitions. As a result, frequency decomposition and scale transformation remain only weakly coupled, limiting targeted regulation of stripe-sensitive frequency information at critical early stages.

Cross-level feature transmission has received considerably less attention from the perspective of stripe degradation. U-shaped destriping networks commonly employ skip connections to recover high-resolution details, and DestripeCycleGAN [6], for example, introduces a Group Fusion Block into skip connections to aggregate multi-scale features and establish long-range dependencies. However, existing cross-level designs mainly regard encoder features as complementary information for reconstruction. Explicitly treating skip connections as propagation paths of residual stripe contamination is rarely reported in infrared destriping networks. In addition, stripe contamination is not distributed uniformly across feature scales: shallow representations preserve relatively complete stripe morphology, whereas deeper features progressively emphasize structural and semantic information. Existing skip-fusion strategies generally do not explicitly model this scale-dependent degradation difference, leaving residual stripe responses in shallow features with a direct route back into the reconstruction process.

Overall, infrared destriping has been extensively studied through nonuniformity correction, single-image prior-based restoration, and deep learning-based approaches. Recent studies have increasingly emphasized directional structures, column-wise characteristics, frequency-domain information, and multi-scale feature interactions, substantially improving stripe suppression and structural preservation. Nevertheless, existing specialized designs mainly exploit these degradation characteristics within particular representations or operations, while the evolution and propagation of stripe-related responses through hierarchical feature transformations remain insufficiently characterized. This limitation calls for more mechanism-oriented designs that explicitly account for how stripe degradation changes across feature representation, scale transformation, and cross-level transmission.

3. Proposed Method

Section 3.1 presents the overall U-shaped architecture. Section 3.2 then introduces the Stripe-Aware Hybrid Module (SHM) and the embedded SSA mechanism. Section 3.3 and Section 3.4 further detail the High-Frequency Gated Sub-band Downsampling module (HFGS-Down) and the Structure-Guided Skip Refinement module (SSR), respectively.

3.1. Network Architecture

The overall architecture of the proposed network is shown in Figure 1. It is a typical four-stage asymmetric U-shaped encoder-decoder that learns the residual mapping from the degraded infrared input to the clean output. The network introduces task-specific processing at feature extraction, scale transformation, and cross-level fusion to accommodate the different manifestations of stripe degradation in hierarchical features.

Specifically, the degraded input image is first projected into a high-dimensional feature space by a $3\times 3$ convolution to obtain shallow features. These features are then enhanced by the Stripe-Aware Hybrid Module (SHM), in which the Stripe-Aware Spatial Attention (SSA) mechanism is embedded, yielding more discriminative shallow representations.

In the encoder, a four-level progressive downsampling structure is adopted for hierarchical feature learning. During the first scale transition, the proposed HFGS-Down module decomposes features into multiple Haar-wavelet sub-bands, adaptively modulates the high-frequency components, and fuses the main and auxiliary paths to reshape stripe-related high-frequency information. The subsequent downsampling stages employ convolutions with a stride of 2 to further extract high-level semantic information.

At the bottleneck, the feature maps reach their minimum spatial resolution and maximum channel dimension. Stacked SHM blocks are then used to transform deep features and aggregate global contextual information.

In the decoder, the spatial resolution is progressively restored stage by stage. At each level, a $1\times 1$ convolution first expands the number of channels, followed by PixelShuffle-based upsampling for efficient reconstruction [21]. Meanwhile, the proposed Structure-Guided Skip Refinement module (SSR) selectively filters the skip features from the encoder. At deeper scales, SSR performs lightweight semantic filtering. At the shallowest scale, an enhanced SSR variant is employed to strengthen stripe-direction modeling through asymmetric directional convolutions, thereby suppressing stripe leakage in high-resolution features. The refined skip features are fused with the decoder features and further processed by SHM to obtain a more accurate reconstruction.

Finally, the high-resolution decoder features are projected back to the single-channel image space through a $3\times 3$ convolution to generate a residual map, which is added to the degraded input image to obtain the final destriped result.

3.2. Stripe-Aware Module Based on Column Statistics

As shown in Figure 1a, SHM is designed as a lightweight stripe-aware hybrid transformation unit and serves as the basic building block of the proposed network. Rather than relying on generic feature mixing alone, SHM integrates local spatial aggregation, gated feature interaction, and stripe-aware selective recalibration into a unified residual formulation. Specifically, the input features are first normalized by Layer Normalization (LN) [22] and expanded to a higher-dimensional intermediate space, where depthwise convolution and the SimpleGate mechanism [23] are used to perform efficient local feature mixing. Since detector-induced vertical stripes are typically associated with persistent response variations across detector columns, column-wise statistics provide informative cues for characterizing stripe-related abnormalities. On this basis, the embedded SSA characterizes the intermediate responses from complementary column-statistical perspectives, enabling stripe-related anomalies to be selectively recalibrated before projection back to the original feature dimension. A second lightweight transformation branch is then applied for complementary feature refinement. In this way, SHM enables efficient feature propagation while continuously enhancing stripe-sensitive representations throughout the encoder, bottleneck, and decoder. Figure 2 illustrates SSA, which consists of three branches: an average branch (AVG), a smooth-maximum branch (SMOOTHMAX), and a differential branch (DIFF).

We first compute the column-wise mean along the height dimension $M_{\mathrm{avg}}\left(b,c,1,w\right)={\displaystyle \frac{1}{H}}{\sum }_{h=1}^{H}X\left(b,c,h,w\right)$. This operation suppresses local random fluctuations and provides a stable background baseline for each column.

To emphasize strong stripe responses that remain coherent along the height direction while containing local variations, we further aggregate features with a learnable-temperature SmoothMax operator $\beta \in \left[{\beta }_{min},{\beta }_{max}\right]$:

$${\alpha }_h={\displaystyle \frac{\exp \left(\beta X_h\right)}{{\sum }_{h^{\prime }}\exp \left(\beta X_{h^{\prime }}\right)}},M_{\mathrm{smoothmax}}={\displaystyle \sum _{h=1}^H}{\alpha }_h{X}_h,$$

(1)

where $X_h$ denotes the feature response at the h-th position in a given channel and column, and ${\alpha }_h$ is the corresponding SmoothMax weight. Compared with hard max pooling, SmoothMax is differentiable and its sharpness can be adjusted continuously, enabling the model to focus on the strongest abnormal response in a column while maintaining stable gradients.

After obtaining the column mean and the smoothed extreme response, we construct their difference term:

$$M_{\mathrm{diff}}=M_{\mathrm{smoothmax}}-M_{\mathrm{avg}}.$$

(2)

This term captures the abnormal deviation of the salient response from the stable baseline, and the final gate is given by

$$G=\sigma \left(\mathrm{Conv}\left(M_{\mathrm{avg}}+M_{\mathrm{smoothmax}}\right).+\lambda ·\mathrm{Conv}\left(M_{\mathrm{diff}}\right)\right)$$

(3)

Here, $\lambda$ is a learnable scaling factor. The three branches correspond to the background baseline, salient response, and abnormal deviation, respectively. The branch-wise visualization and ablation results presented later also show that combining all three statistics is consistently superior to using any single branch alone.

3.3. High-Frequency Gated Sub-Band Downsampling

Shallow features retain relatively explicit stripe-related directional and frequency responses, which may become mixed with genuine structures during downsampling. If only strided convolution is used, stripe-related high-frequency components and genuine structures are mixed during downsampling. If only wavelet decomposition is used, directional sub-bands can be preserved explicitly, but semantic interaction remains limited. HFGS-Down is therefore applied at stage 0 and adopts a dual-path design: one path uses the DWT to preserve directional frequency information, while the other path uses strided convolution to complement conventional spatial feature aggregation. More importantly, the module uses the frequency state of the input features to regulate both high-frequency responses and the subsequent fusion of the two paths, rather than simply employing wavelet decomposition as an additional representation.

Unlike existing dual-path sampling schemes, the core of HFGS-Down lies in the joint modeling of the high-frequency family. Given an input feature map $X\in {\mathbb{R}}^{B\times C\times H\times W}$, where B, C, H, and W denote the batch size, number of channels, feature height, and feature width, respectively, the Haar discrete wavelet transform decomposes it into four directional sub-bands:

$$\left(F_{LL},F_{HL},F_{LH},F_{HH}\right)=\mathrm{DWT}\left(X\right).$$

(4)

Here, $F_{LL}$ denotes the low-frequency structural body, while $F_{HL}$, $F_{LH}$, and $F_{HH}$ jointly form the high-frequency family. Stripe-related disturbances are not assumed to be confined to a single directional sub-band, since the main stripe response, local texture coupling, and edge interactions in the learned feature space may be jointly distributed across the high-frequency family. Therefore, an effective strategy should not remain at the level of single-sub-band local repair, but should make a more robust judgment based on the overall state of the high-frequency family. To characterize this overall frequency state, the high-to-low frequency response ratio is defined as follows:

$$S={\displaystyle \frac{E_{HL}+E_{LH}+E_{HH}}{E_{LL}}},$$

(5)

where $E_{HL}$, $E_{LH}$, $E_{HH}$, and $E_{LL}$ denote the mean absolute response magnitudes of the corresponding sub-bands, respectively. Based on this statistic, a shared gating function is constructed as

$$g=\sigma \left(k\right(b-log(S+\epsilon )\left)\right),$$

(6)

where k and b are learnable scaling and bias parameters, respectively, $\sigma (·)$ denotes the Sigmoid function, and $\epsilon$ is a small positive constant for numerical stability. This gate does not operate on one local high-frequency component in isolation. Instead, it jointly modulates the entire high-frequency family according to the overall high-frequency disturbance level relative to the low-frequency body.

The same gating coefficient g is subsequently shared by the three high-frequency sub-bands $F_{HL}$, $F_{LH}$, and $F_{HH}$, whereas the $F_{LL}$ branch remains relatively stable. The rationale behind this design is that stripe contamination in deep feature space is more likely to appear as joint activation of the high-frequency family than to be strictly confined to a single individual sub-band. Hence, compared with single-sub-band gating, a unified yet moderate joint regulation over the high-frequency family is more robust.

To allow the directional-sub-band path and the convolutional path to cooperate adaptively according to the frequency state of the input, adaptive fusion weights are further derived from the same response statistic:

$$\alpha =\psi \left(S\right),0<\alpha <1.$$

(7)

The final output can then be written as follows:

$$Y=2\left((1-\alpha )Y_{\mathrm{conv}}+\alpha Y_{\mathrm{dwt}}\right),$$

(8)

where $Y_{\mathrm{conv}}$ and $Y_{\mathrm{dwt}}$ denote the outputs of the convolutional and wavelet branches, respectively, and $\psi (·)$ denotes a bounded learnable nonlinear mapping that converts the log-transformed frequency-state statistic into the fusion weight $\alpha$ through affine modulation and a Tanh activation. This design means that the same frequency-state variable S in HFGS-Down controls two behaviors simultaneously: joint reshaping of the high-frequency family and adaptive fusion between the two downsampling paths.

3.4. Heterogeneous Stripe-Leakage Suppression for Hierarchical Skip Connections

In a U-shaped encoder-decoder, skip connections help recover high-resolution details, but they may also feed incompletely purified stripe responses back into the decoder, thereby weakening the destriping effect achieved by earlier stages. For infrared destriping, such leakage does not behave identically across scales. The highest-resolution skip preserves the most complete stripe morphology and has the most direct influence on the output, making it the most dangerous shortcut for stripe leakage. By contrast, deeper skips mainly carry coarse semantic and structural information, and excessively strong local gating at these scales may damage normal semantic fusion. Accordingly, the same gating strategy is not used across all skip levels. Instead, as shown in Figure 1c, a hierarchical heterogeneous SSR design is adopted: deeper skips are refined by a lightweight SSR-Lite module, while the shallowest skip is handled by an enhanced SSR block with stronger local directional suppression.

For deeper skips, SSR-Lite adopts a lightweight gating strategy. Given the upsampled decoder feature g and the encoder skip feature x, the two streams are first aligned and fused, after which a gate is generated as

$${\mathrm{A}}_{\mathrm{lite}}=\sigma \left(\varphi \left({\mathrm{W}}_g\left(g\right)+{\mathrm{W}}_x\left(x\right)\right)\right),$$

(9)

where $W_g$ and $W_x$ represent the alignment mappings of the two inputs, and $\varphi$ represents a nonlinear transformation. This design follows the spirit of lightweight gated fusion [24], but is further adapted here for semantic filtering in deep skips. It suppresses cross-level stripe leakage at low cost while preserving semantic continuity as much as possible.

In contrast, the highest-resolution shallow skip requires explicit handling of stripe-direction characteristics. For this reason, an enhanced SSR is adopted at the shallow stage. The two inputs are first aligned and fused, as shown below:

$$u=\varphi \left({\mathrm{W}}_g\left(g\right)+{\mathrm{W}}_x\left(x\right)\right).$$

(10)

On the fused feature u, two asymmetric depthwise-convolution branches are introduced in parallel, one vertical and one horizontal, and the final gate is produced from the direction-enhanced representation using the following equation:

$$v={\mathrm{DWConv}}_{11\times 1}\left(u\right)+{\mathrm{DWConv}}_{1\times 7}\left(u\right),$$

(11)

$$A=\sigma \left(\mathrm{Conv}\left(v\right)\right).$$

(12)

Here, the $11\times 1$ depthwise convolution captures stripe continuity along the height direction, while the $1\times 7$ depthwise convolution senses inter-column variation and horizontal abnormality. This gate selectively modulates the shallow skip features, so that real edges and textures are preserved while shallow stripe leakage is more effectively suppressed.

The essence of SSR is not simply to attach a gate to the skip pathway, but to apply heterogeneous control according to the information carried by skips at different scales. Deeper skips use lightweight semantic filtering to avoid excessive intervention, whereas the shallowest skip uses the enhanced SSR to perform local directional blocking and cut off the most dangerous stripe shortcut. This yields a hierarchical cooperative strategy of strong shallow suppression and gentle deep filtering, which more effectively reduces stripe leakage during decoding.

4. Experimental Results and Discussion

To comprehensively validate the proposed method for infrared image destriping, the experiments are organized into five aspects: simulated-image restoration, real-scene evaluation, ablation and mechanism analysis, computational efficiency and scalability, and downstream application. First, on synthetic data with available references, PSNR and SSIM are used to quantitatively compare the proposed method with traditional approaches, deep models, and recent representative infrared destriping methods. Second, real-scene validation is conducted on self-acquired infrared data covering building and sky–aircraft scenes. Since paired clean references are unavailable for these real observations, qualitative comparison, spatial- and frequency-domain analysis, and no-reference image-quality metrics are jointly employed to assess restoration behavior. Third, system-level ablations, module-level comparisons, and mechanism visualizations are performed to verify the contribution of SSA, HFGS-Down, and SSR. Fourth, inference time, GPU memory consumption, and multi-resolution scalability are evaluated to examine computational practicality. Finally, downstream transferability is further assessed on a real noisy infrared small-target dataset using multiple destriping preprocessing methods and representative IRSTD detectors.

4.1. Experiment Settings

SFPRNet is compared with nine state-of-the-art infrared destriping methods, including three traditional approaches, ADOM [25], GF [26], and MIRE [27], as well as six advanced deep learning methods, SNRCNN [28], DLS-NUC [29], SNRWDNN [12], NAFNet [23], DSCGAN [6], and ASCNet [7]. For a fair comparison, all learning-based methods are retrained using the same datasets and training settings as SFPRNet.

Dataset Construction. Because large-scale paired infrared destriping datasets with clean reference images are scarce, especially for remote-sensing scenes, multiple public infrared datasets are combined to increase scene diversity and improve model generalization. The resulting multi-source and multi-scene dataset covers representative content, including buildings, vehicles, pedestrians, vegetation, and cloud imagery. Simulated stripe degradations are used for paired training and evaluation, while self-acquired real infrared data are used for real-scene validation.

Specifically, the training and test images are drawn from five data sources: IR700 [30], ESPOL [31], DLS-NUC [29], ICSRN [20], and FLIR [32]. IR700 contains 700 images, including 600 training images and 100 test images. ESPOL contains 101 images, including 91 training images and 10 test images. DLS-NUC contains 100 images, including 93 training images and 7 test images. A total of 48 infrared cloud images are selected from ICSRN, including 42 images for training and 6 images for testing. FLIR contains 50 images, including 45 training images and 5 test images. In total, the training set contains 871 images, and the test set contains 128 images.

For real-scene evaluation, additional infrared images are acquired using a self-developed cooled short-wave infrared (SWIR) avalanche photodiode (APD) camera with an image resolution of $640\times 512$ pixels. The acquired data cover two representative scene types: building scenes containing dense structural textures and sky–aircraft scenes containing relatively smooth backgrounds and airborne targets. Bad-pixel correction is performed on the acquired images before destriping evaluation, while the visible column-wise nonuniformity is retained for real-image validation. Since paired clean references are unavailable for these real observations, the self-acquired images are used only for real-scene evaluation rather than supervised training.

Evaluation Metrics. To quantitatively evaluate infrared destriping performance, two widely used full-reference image-quality metrics, peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) [33], are adopted. PSNR measures pixel-level fidelity and usually increases as the restored image becomes closer to the reference. SSIM measures structural similarity, and values closer to 1 indicate better preservation of scene details and structural information.

Training Degradation Model. To keep the degradation model consistent with prior work and to improve the modeling of complex column-wise nonuniformity, we use a widely adopted cubic degradation model during training to simulate column response mismatch in infrared detectors [29,34]. Let the clean infrared image be denoted by $I_C$ and the degraded image by $I_D$. The training degradation is defined as follows:

$$I_D=I_C+A_1+A_2{I}_C+A_3{I}_C^2+A_4{I}_C^3,$$

(13)

where $A_1$, $A_2$, $A_3$, and $A_4$ denote random coefficient vectors that are expanded along the column direction and broadcast to the entire image. They correspond to the additive bias term, the first-order response mismatch term, the second-order nonlinear term, and the third-order nonlinear term, respectively.

Training Settings. Using the above training images, 4355 paired single-channel image patches of size $256\times 256$ are obtained after preprocessing. During training, random $128\times 128$ patches are cropped from these samples, and random horizontal and vertical flipping is used for data augmentation. The batch size is set to 128. The model is optimized using the AdamW optimizer. The initial learning rate is set to $1\times {10}^{-3}$ and gradually decayed to $1\times {10}^{-6}$ with a cosine annealing schedule. The total training process runs for about 500 epochs, including approximately 30 warm-up epochs. Model training is conducted on two NVIDIA V100 GPUs, with a total training time of approximately 2 h.

4.2. Simulated Image Destriping

4.2.1. Quantitative Results

Three representative stripe-degradation models are considered to evaluate the model performance.

• Gaussian: Stripe noise following a Gaussian distribution with noise level $\sigma \in [0,0.10]$.
• Mixed: On top of the Gaussian stripe noise, additional Gaussian random noise with noise level $\sigma \in [0,0.05]$ is added.
• Wide-period: Considering that practical infrared detectors often suffer from wide-stripe column noise caused by mismatch in readout channels, column amplifiers, or segmented circuits, we further design a wide-period stripe model in the testing stage to better reflect engineering scenarios. Specifically, the image is divided along the column direction into stripe blocks of width W. For the k-th stripe block, additive and multiplicative perturbations are jointly applied, and the degradation is defined as

  $$I_D^{\left(k\right)}=\left(1+G_k\right)I_C^{\left(k\right)}+B_k+J_k,$$

  (14)

  where $G_k$ denotes the gain perturbation with $\sigma \in [0,0.10]$, $B_k$ denotes the bias perturbation with $\sigma \in [0,0.10]$, and $J_k$ denotes the intra-block pixel-level jitter term with $\sigma \in [0,0.01]$.

The quantitative results are reported in

Table 1. Quantitative comparison of different methods on the test sets in terms of PSNR and SSIM.

Category	Index	ADOM	GF	MIRE	SNRCNN	DLS-NUC	SNRWDNN	NAFNet	DSCGAN	ASCNet	Ours
IR700-100
Gaussian	PSNR↑	30.0083	34.8742	32.8174	33.7442	38.1721	38.0898	39.6361	41.3426	42.1435	42.6997
	SSIM↑	0.8547	0.9488	0.9373	0.9348	0.9778	0.9773	0.9848	0.9892	0.9902	0.9920
Wide	PSNR↑	30.6482	35.8570	33.7246	34.2055	37.9809	38.4726	40.2509	41.7473	42.0397	42.8232
	SSIM↑	0.8766	0.9604	0.9527	0.9399	0.9708	0.9770	0.9856	0.9891	0.9887	0.9919
Mixed	PSNR↑	26.9224	30.1896	29.3211	30.2246	31.9537	31.9174	32.6617	33.2814	33.2906	33.5054
	SSIM↑	0.6183	0.7401	0.7199	0.7576	0.7689	0.7706	0.7819	0.7840	0.7788	0.7843
ESPOL-10
Gaussian	PSNR↑	31.7440	34.8661	33.2875	34.6076	38.5380	38.2140	39.7665	40.9548	42.0727	42.0505
	SSIM↑	0.8482	0.9391	0.9213	0.9271	0.9728	0.9682	0.9802	0.9815	0.9868	0.9869
Wide	PSNR↑	33.7932	39.3867	36.4989	37.7068	40.7759	40.5679	42.2772	42.8157	43.2341	43.7391
	SSIM↑	0.9010	0.9758	0.9649	0.9551	0.9773	0.9747	0.9848	0.9860	0.9878	0.9905
Mixed	PSNR↑	29.3728	32.1810	31.1669	32.4177	34.3047	34.2190	35.1212	35.8523	35.7840	35.8576
	SSIM↑	0.6807	0.7958	0.7750	0.8074	0.8302	0.8287	0.8455	0.8484	0.8402	0.8416
DLS-NUC-7
Gaussian	PSNR↑	32.5929	36.7779	27.8496	36.0688	39.7410	39.8097	41.1839	42.2749	42.5898	43.1999
	SSIM↑	0.8658	0.9574	0.9119	0.9434	0.9814	0.9786	0.9849	0.9893	0.9891	0.9900
Wide	PSNR↑	33.3475	39.3535	30.6004	37.5067	39.7598	40.6534	42.8129	43.2991	43.3380	43.7635
	SSIM↑	0.8945	0.9753	0.9377	0.9542	0.9700	0.9781	0.9886	0.9895	0.9899	0.9907
Mixed	PSNR↑	28.6296	32.0168	26.9560	32.2658	33.4090	33.5140	33.9538	34.1541	34.1894	34.3356
	SSIM↑	0.5839	0.7006	0.6539	0.7245	0.7251	0.7250	0.7331	0.7333	0.7280	0.7333
ICSRN-6
Gaussian	PSNR↑	28.8455	36.1407	20.3065	32.9996	37.3004	37.4397	37.9297	38.5954	38.3843	38.6429
	SSIM↑	0.8565	0.9724	0.8361	0.9391	0.9836	0.9850	0.9900	0.9935	0.9743	0.9946
Wide	PSNR↑	28.0229	33.8100	21.6958	30.6477	33.8766	34.8384	35.6340	37.6272	35.7142	38.0046
	SSIM↑	0.8376	0.9625	0.6896	0.9016	0.9598	0.9747	0.9846	0.9905	0.9597	0.9919
Mixed	PSNR↑	26.4836	29.8587	19.2597	29.2005	30.8913	30.9439	31.3420	31.1852	31.0959	31.2243
	SSIM↑	0.6760	0.7898	0.8306	0.7825	0.8059	0.8093	0.8173	0.8104	0.8027	0.8195
FLIR-5
Gaussian	PSNR↑	31.3692	36.0267	29.0257	34.7769	38.2809	38.3639	39.2340	40.2687	41.0845	41.1531
	SSIM↑	0.8784	0.9626	0.9440	0.9508	0.9849	0.9854	0.9900	0.9926	0.9941	0.9949
Wide	PSNR↑	31.7260	36.9682	28.6674	35.6216	39.1034	39.1463	40.0750	41.0609	41.0922	41.3658
	SSIM↑	0.8808	0.9627	0.9434	0.9544	0.9838	0.9849	0.9893	0.9915	0.9909	0.9937
Mixed	PSNR↑	28.7537	31.0877	28.0160	31.2624	32.6887	32.6225	33.1430	33.3958	33.4278	33.6637
	SSIM↑	0.6705	0.7643	0.7488	0.7815	0.7916	0.7931	0.8033	0.8043	0.7969	0.8045

Note: ↑ indicates that a higher value is better. Bold and underlined metric values indicate the best and second-best results, respectively.

, where the best and second-best results are marked in boldface and underlined, respectively. SFPRNet achieves the best or second-best PSNR/SSIM values in most test scenarios. In particular, on the relatively larger IR700-100 test set, the proposed method shows a consistently stronger advantage under all three degradation settings.

4.2.2. Visual Performance

The qualitative destriping results are shown in Figure 3, Figure 4 and Figure 5, where the three groups of test images correspond to Gaussian noise, wide-period noise, and mixed noise, respectively. It can be observed that traditional methods still leave noticeable stripe residuals in some regions and tend to oversmooth real structures. Although several deep learning methods can suppress stripes to some extent, they may still suffer from detail blurring or local structural distortion in areas with complex textures. By contrast, the proposed method achieves better overall restoration quality under all three degradation settings. In Figure 3, the background window contours recovered by our method are closer to the ground truth. In Figure 4, stripe contamination around the window frames and staircase regions is more thoroughly suppressed. In Figure 5, the model removes noise while better preserving the texture details on the building surface. These results indicate that SFPRNet achieves a more favorable balance between stripe suppression strength and structural fidelity.

4.3. Real Image Destriping

To further assess the generalization ability of the model in real scenes, we conduct comparative experiments on real images captured by a custom-built infrared camera, as shown in Figure 6. Although some deep learning methods can alleviate stripe interference, they still suffer from texture loss or structural oversmoothing around building surfaces and brick edges. By contrast, the proposed method more effectively suppresses column-direction stripes while preserving the brick texture and boundary contours of buildings, resulting in more stable visual restoration on real infrared data.

To further examine the restoration behavior on self-acquired real infrared remote-sensing data, we conduct spatial- and frequency-domain analyses. As shown in Figure 7, the degraded image exhibits pronounced column-wise fluctuations and distinct stripe-related peaks in the horizontal frequency spectrum. After processing by SFPRNet, the abnormal column responses and spectral peaks are substantially suppressed, while the overall scene structure and signal variation are well maintained.

Since clean reference images are unavailable for the self-acquired real infrared remote-sensing data, we further introduce no-reference evaluation to complement the qualitative comparison and the spatial- and frequency-domain analyses. The stripe-aware no-reference metric $\rho$ is used to measure residual column-wise nonuniformity, where a smaller value indicates weaker stripe artifacts. In addition, the natural image quality evaluator (NIQE) [35] is adopted to assess the overall perceptual quality without reference images. As reported in

Table 2. No-reference quantitative comparison on real infrared images.

Metric	ADOM	GF	SNRCNN	DLS-NUC	SNRWDNN	NAFNet	DSCGAN	ASCNet	Ours
$\rho$↓	0.1029	0.0817	0.0726	0.0787	0.0785	0.0784	0.0815	0.0792	0.0776
NIQE↓	8.7811	5.1341	6.5920	5.9685	5.7795	5.1743	5.1067	5.2968	4.9063

Note: ↓ indicates that a lower value is better. Bold and underlined values indicate the best and second-best results, respectively.

, SFPRNet achieves the best or second-best results on the no-reference metrics, indicating a favorable balance between stripe suppression and real-scene image quality. Together with the qualitative comparisons and spatial- and frequency-domain analyses, these results provide complementary evidence of the effectiveness of SFPRNet on real infrared observations without paired clean references.

4.4. Ablation Studies and Mechanism Analysis

To fairly evaluate the individual contribution of each module, all subsequent ablation experiments are retrained under the same configuration, and the average PSNR and SSIM over five synthetic Gaussian-stripe test subsets are reported.

Table 3. Step-wise system-level ablation of the proposed mismatch rectification chain.

Group	Baseline	SSA	HFGS-Down	SSR	PSNR↑/SSIM↑
1	✓	✕	✕	✕	39.6352/0.9850
2	✓	✓	✕	✕	41.9334/0.9900
3	✓	✓	✓	✕	42.3063/0.9914
4	✓	✓	✕	✓	42.2656/0.9907
5	✓	✓	✓	✓	42.5487/0.9915

Note: ✓ indicates that the corresponding module is included, whereas ✕ indicates that it is excluded. ↑ indicates that a higher value is better. Bold values indicate the best performance.

presents the step-wise system-level ablation results of the proposed framework. As the three modules are introduced progressively, the performance improves consistently, showing that the corresponding task-specific designs provide stable gains at different stages of the restoration process.

To further examine the robustness of these improvements to random initialization, additional training runs with different random seeds are conducted for Groups 2–5, while the original baseline result (Group 1) is retained as a reference. As shown in Figure 8, repeated runs of each configuration are tightly clustered in the PSNR–SSIM space, with PSNR standard deviations below 0.05 dB. The mean performance trend remains consistent with the step-wise ablation results, and the final configuration maintains the best overall performance across different random seeds. This indicates that the observed improvements are not attributable to a single favorable initialization.

4.4.1. Stripe-Aware Spatial Attention

To verify the effectiveness of the proposed SmoothMax-SSA for infrared destriping, we keep the baseline unchanged and only replace the feature-selection mechanism inside the encoder block. The proposed module is compared with several representative attention or correction modules, including CBAM [36], EA [37], SCA [23], and RCSSC [7]. These modules cover several major comparison dimensions, ranging from generic two-dimensional saliency modeling and lightweight efficient self-attention to stripe-oriented column-aware correction.

The quantitative results in

Table 4. Quantitative comparison of different attention modules.

Method	PSNR↑	SSIM↑	Params (M)↓	FLOPs (G)↓
CBAM	39.7568	0.9857	3.303	10.779
EA	39.9951	0.9864	4.807	16.364
SCA	39.6352	0.9850	3.640	10.745
RCSSC	41.3384	0.9881	4.119	18.168
SSA	41.9344	0.9900	3.644	10.797

Note: ↑ indicates that a higher value is better, whereas ↓ indicates that a lower value is better. Bold values indicate the best result in each column.

, with the best values highlighted in bold, show that SSA achieves the best restoration performance while keeping the parameter count and computational cost relatively low. This suggests that the proposed column-statistics modeling strategy is more suitable for infrared destriping than generic two-dimensional attention or general feature correction modules. As further shown in Figure 9, SSA restores fine structures such as window boundaries and wooden poles more faithfully.

To verify the roles of the three branches in SSA, namely Avg, SMOOTHMAX, and DIFF, we conduct a dedicated ablation study and visualize their responses on typical stripe-contaminated samples using a trained SSA model. A sequential colormap is used to display response intensity, allowing a direct comparison of how different statistical mechanisms model column-direction information from the same input.

As shown in

Table 5. Ablation study of the internal statistical branches in SSA.

U-Net	AVG	SMOOTHMAX	DIFF	PSNR↑/SSIM↑
✓	✕	✕	✕	39.6352/0.9850
✓	✓	✕	✕	41.7543/0.9897
✓	✓	✓	✕	41.8895/0.9900
✓	✓	✓	✓	41.9344/0.9900

Note: ✓ indicates that the corresponding component is included, whereas ✕ indicates that it is excluded. ↑ indicates that a higher value is better. Bold values indicate the best performance.

, performance improves steadily as the AVG, SMOOTHMAX, and DIFF branches are introduced step by step, indicating clear complementarity among the three statistical cues. The visualizations in Figure 10 further reveal their different roles. The AVG branch responds over a broader area and mainly captures the stable background baseline along the column direction. The SMOOTHMAX branch is more concentrated on local salient responses and is more sensitive to abnormal peaks. The DIFF branch further characterizes the deviation of salient responses from the mean statistics and forms clearer column-wise patterns in stripe regions. These observations show that SSA does not rely on a single statistic; instead, it jointly models global baselines, salient responses, and abnormal deviations to strengthen the perception and suppression of vertical stripe noise.

4.4.2. HFGS-Down

The Haar wavelet is adopted in HFGS-Down because its short-support filter bank provides highly localized decomposition, while its high-pass component captures local differences between adjacent responses. These characteristics are consistent with the localized inter-column response variations associated with infrared stripe nonuniformity, making Haar a reasonable choice for the first scale transition. To examine the influence of the wavelet basis, Haar, db2, sym5, and coif1 are compared under the same network and training settings.

As shown in

Table 6. Quantitative comparison of different wavelet bases in HFGS-Down.

Wavelet Basis	Haar	db2	sym5	coif1
PSNR↑	42.2676	42.0434	42.0675	42.1090
SSIM↑	0.9914	0.9911	0.9910	0.9910

Note: ↑ indicates that a higher value is better. Bold values indicate the best result.

, Haar achieves the highest PSNR and SSIM among the four tested wavelet bases, providing experimental support for its use in HFGS-Down. Based on this setting, the stage-0 downsampling strategy is further examined while keeping the rest of SFPRNet unchanged.

Conventional strided-convolution downsampling, Haar-DWT-only downsampling, Conv + DWT, which simply combines convolution and wavelet branches without the proposed high-frequency gating, and the proposed HFGS-Down are compared. The purpose of this experiment is not simply to determine whether wavelets are used, but to examine whether the network can perform truly effective targeted regulation over the high-frequency family once directional sub-bands have been explicitly unfolded.

As reported in

Table 7. Quantitative comparison of different downsampling strategies.

Method	Conv-Only	DWT-Only	Conv + DWT	HFGS-Down
PSNR↑	41.9344	42.1165	42.1524	42.2784
SSIM↑	0.9900	0.9910	0.9911	0.9914

Note: ↑ indicates that a higher value is better. Bold values indicate the best result.

, the proposed HFGS-Down achieves the best performance among all downsampling strategies, indicating that explicit preservation of directional sub-bands together with dual-path fusion in the first scale transition benefits the representation of stripe-related high-frequency information. Furthermore,

Table 8. Ablation on sub-band gating strategies of HFGS-Down.

Method	PSNR↑	SSIM↑
HL/LL	42.1183	0.9913
(HL + LH + HH)/LL	42.2784	0.9914

Note: ↑ indicates that a higher value is better. Bold values indicate the best result.

shows that gating the high-frequency family jointly yields better performance than gating a single sub-band alone. This suggests that, in deep feature space, stripe-related disturbances are not strictly confined to one directional sub-band, but tend to appear as joint activation across the high-frequency family. Therefore, shared regulation over the entire high-frequency family is more robust than single-sub-band modulation.

An external high-frequency indicator that is decoupled from the gating formula itself is further introduced to independently measure the level of high-frequency activity in each input sample. Given an input image I, we compute first-order differences in the horizontal and vertical directions, respectively:

$${\partial }_{x}I(i,j)=I(i,j+1)-I(i,j),{\partial }_{y}I(i,j)=I(i+1,j)-I(i,j).$$

(15)

The average of their absolute values is then used as the external high-frequency indicator $H_{ext}$:

$$H_{\mathrm{ext}}={\displaystyle \frac{1}{2}}\left({\displaystyle \frac{1}{N_x}}\sum \left|{\partial }_{x}I\right|+{\displaystyle \frac{1}{N_y}}\sum \left|{\partial }_{y}I\right|\right),$$

(16)

where $N_x$ and $N_y$ denote the numbers of valid horizontal and vertical difference samples, respectively.

As shown in Figure 11, the fusion coefficient increases noticeably with the external high-frequency indicator, implying that the model adaptively increases its reliance on the wavelet branch when the input contains stronger high-frequency variations. The grouped statistics further verify this trend: although the convolutional branch remains dominant overall, the contribution of the wavelet branch increases as the high-frequency activity of the sample becomes stronger, which helps preserve stripe-related high-frequency discriminative cues.

4.4.3. Structure-Guided Skip Refinement

The effectiveness of SSR is further evaluated by comparing four skip-fusion strategies: (a) direct summation without gating; (b) applying SSR-Lite to all skip connections; (c) applying the enhanced SSR to all skip connections; and (d) the proposed hierarchical SSR design, where the shallow skip uses SSR and deeper skips use SSR-Lite.

The quantitative results in

Table 9. Quantitative comparison of different skip gating strategies.

Method	PSNR↑	SSIM↑	Params (M)↓
Baseline	41.9344	0.9900	3.644
SSR-Lite	42.1798	0.9903	3.802
SSR	42.2516	0.9904	3.808
SSR + SSR-Lite	42.2331	0.9907	3.803

Note: ↑ indicates that a higher value is better, whereas ↓ indicates that a lower value is better. Bold values indicate the best result in each column.

, with the best values highlighted in bold, show that introducing skip gating consistently improves performance, confirming that cross-level feature filtering is necessary for infrared destriping. The proposed hierarchical heterogeneous strategy achieves a more balanced improvement in both PSNR and SSIM, with a particularly noticeable gain in SSIM. This indicates that shallow and deep features should be handled with different gating intensities: shallow features require explicit suppression of stripe residuals, whereas deep features benefit more from gentle semantic filtering so that normal structural transmission is not over-constrained. To further analyze the internal behavior of SSR, the gating process on the shallowest skip connection is visualized, as shown in Figure 12.

An infrared image heavily contaminated by stripes is selected as the input. From the learned gating mask in Figure 12c, clear continuous column-wise stripe structures can be observed, indicating that the module is able to target stripe-contaminated regions selectively. A comparison between the shallow feature maps before and after gating in Figure 12d,e shows that the originally scattered and disordered high-frequency fragments are markedly reduced, while more coherent vertical structures remain. The column-wise statistical responses in Figure 12f further show that the overall response energy decreases after gating, but the amount of attenuation is not uniform across columns. This demonstrates that SSR does not simply suppress all columns equally; instead, it performs differentiated filtering according to the contamination level of each column. Overall, SSR can selectively weaken shallow stripe residuals while preserving useful structural information, thereby mitigating stripe leakage during decoding.

4.5. Computational Efficiency and Scalability

To evaluate the computational efficiency and scalability of SFPRNet, inference experiments are conducted on a single NVIDIA V100 GPU using a batch size of 1 and FP32 precision. All images are processed by direct full-image inference without patch-based tiling. After 30 warm-up runs, the average inference time is measured over 100 forward passes with CUDA synchronization, while the peak GPU memory consumption is recorded during inference. The corresponding frame rate is calculated from the average inference time.

As shown in

Table 10. Computational efficiency and scalability of SFPRNet at different input resolutions.

Input Resolution	Avg. Inference Time (ms)	Peak GPU Memory (MB)	FPS
$256\times 256$	18.6596	191.25	53.5917
$512\times 512$	25.6298	381.00	39.0171
$640\times 512$	30.9626	444.25	32.2970
$1024\times 1024$	94.7764	1140.00	10.5512

, the computational cost increases with input resolution, while SFPRNet maintains favorable scalability across the tested resolutions. Under the tested hardware configuration, the model provides real-time or near-real-time processing capability at moderate input resolutions. For higher-resolution inputs, the increased computational cost indicates that further acceleration may be required to meet real-time processing requirements.

4.6. Task-Driven Evaluation on IRSTD

The practical value of high-quality destriping lies in alleviating the performance degradation of downstream high-level vision tasks. To evaluate the influence of different destriping preprocessing methods on infrared small-target detection (IRSTD), three representative detectors, UIU-Net [38], RDIAN [39], and DNA-Net [40], are evaluated on the noisy IRSTDID-800 dataset [41] using their publicly available pre-trained weights. The evaluation metrics include detection probability ($P_d$), false-alarm rate ($F_a$), intersection over union (IoU), and F-measure.

As shown in

Table 11. Quantitative comparison of downstream infrared small-target detection performance with different destriping preprocessing methods.

Method	UIU-Net	RDIAN	DNA-Net
	${\mathit{P}}_{\mathit{d}}\uparrow /{\mathit{F}}_{\mathit{a}}\downarrow /\mathbf{IoU}\uparrow /$F-Measure↑	${\mathit{P}}_{\mathit{d}}\uparrow /{\mathit{F}}_{\mathit{a}}\downarrow /\mathbf{IoU}\uparrow /$F-Measure↑	${\mathit{P}}_{\mathit{d}}\uparrow /{\mathit{F}}_{\mathit{a}}\downarrow /\mathbf{IoU}\uparrow /$F-Measure↑
Noisy	72.6010/0.5078/50.3540/66.9806	73.0380/0.0687/34.4587/51.2554	88.1980/1.1597/60.4434/75.3455
MIRE	74.7475/0.4883/50.2041/66.8478	70.5063/0.0534/40.0875/57.2321	89.9746/2.1973/58.0863/73.4868
ASCNet	82.9949/0.5469/52.6240/68.9590	76.9036/0.0687/38.3286/55.4167	92.8934/2.2888/60.6533/75.5083
SFPRNet	85.3535/0.5469/52.8501/69.1529	79.4768/0.0687/38.3701/55.4601	94.6701/2.1973/60.6827/75.5311

Note: ↑ indicates that a higher value is better, whereas ↓ indicates that a lower value is better. Bold values indicate the best $P_d$ result for each detector.

, SFPRNet consistently achieves the highest detection probability ($P_d$) among all preprocessing conditions for all three detectors. In particular, for UIU-Net, $P_d$ increases from 72.60% on the original noisy images to 85.35% after SFPRNet preprocessing, corresponding to an improvement of 12.75 percentage points. Improvements are also observed for RDIAN and DNA-Net, indicating that the restored images provide more favorable inputs for detecting infrared small targets. The IoU and F-measure results further show competitive downstream performance overall. The different improvement magnitudes across the three detectors suggest that the benefit of destriping is related to their different sensitivities to structured background interference and their inherent robustness to noisy inputs.

Figure 13 further provides representative visual examples. After stripe interference is suppressed, weak target responses become more distinguishable from the structured background, reducing the interference of stripe artifacts with target perception. These observations are consistent with the improvement in detection probability and demonstrate the practical value of SFPRNet as a preprocessing method for downstream infrared small-target detection.

5. Discussion and Limitations

Although SFPRNet achieves favorable restoration performance and computational efficiency, several issues remain for practical infrared imaging applications. The current method mainly focuses on stripe-dominant degradation, whereas real infrared imaging systems may simultaneously suffer from random noise, stripe patterns with varying orientations, more complex fixed-pattern artifacts, optical distortion, and other coupled degradations. Variations caused by detector aging, temperature drift, and different operating conditions may further change the characteristics and distribution of stripe noise. Beyond artifact removal, practical infrared imaging also involves broader image-quality challenges, such as limited dynamic range and spatial resolution, which are not explicitly addressed in the current framework. In addition, although the computational experiments demonstrate favorable scalability, deployment on resource-constrained infrared imaging platforms still requires further consideration of inference efficiency, memory consumption, and hardware-oriented optimization. Future work will therefore investigate more flexible mixed-degradation modeling, adaptation to diverse sensors and imaging conditions, joint restoration of multiple image-quality degradations, and efficient deployment on practical infrared imaging systems.

6. Conclusions

This paper presents SFPRNet, a spatio-frequency synergistic progressive restoration network for infrared image destriping. Rather than treating stripe degradation as a generic image-restoration problem, SFPRNet performs task-specific modeling according to the statistical response characteristics, frequency behavior during scale transformation, and cross-level propagation of detector-induced stripes. Stripe-Aware Spatial Attention enhances degradation-specific column-statistical characterization, High-Frequency Gated Sub-band Downsampling adaptively regulates the first scale transformation according to the frequency state of shallow features, and Structure-Guided Skip Refinement suppresses residual stripe transmission through scale-dependent cross-level refinement. These designs improve stripe discrimination and suppression while maintaining structural information throughout hierarchical restoration.

Extensive experiments on synthetic and real infrared data demonstrate that SFPRNet achieves superior or competitive restoration performance across different scenes and degradation settings, while maintaining a favorable balance between stripe suppression and structural detail preservation. The restored images also provide more reliable inputs for downstream infrared small-target detection, further demonstrating the practical value of SFPRNet for infrared image enhancement and subsequent scene analysis.