Frequency-Spatial Gated Network (FSG-Net)

• Frequency-Spatial Gated Networks are architectures that combine spatial and frequency-aware feature modeling with adaptive gating.
• They employ dual-path processing and explicit low/high frequency decomposition to improve tasks such as infrared detection and super-resolution.
• Adaptive fusion methods like depthwise blur, Laplacian decomposition, and cross-domain attention underpin enhanced performance across various applications.

Frequency-Spatial Synergistic Gated Network (FSG-Net) denotes an architectural pattern in which spatial-domain feature modeling and frequency-aware feature modeling are coupled through adaptive gates, attention-like selectors, or residual modulation. The term is not standardized across the literature. The exact name FSGNet appears in infrared small target detection as “Frequency-Aware and Semantic Guidance Network,” not as “Frequency-Spatial Synergistic Gated Network,” while several closely related models instantiate the same underlying idea under different names. This suggests that FSG-Net is best understood as a family of frequency-spatial gated designs rather than a single universally fixed blueprint (Zhang et al., 26 Mar 2026).

1. Terminology and scope

The closest directly named model is FSGNet for infrared small target detection, but the broader design pattern also includes transformer, CNN, dual-encoder, and recurrent-free predictor variants that explicitly separate or approximate frequency content, preserve spatial structure, and fuse the two with learnable control.

Model	Task	Frequency-spatial/gated mechanism
FSGNet (Zhang et al., 26 Mar 2026)	Infrared small target detection	FFT-based skip refinement, semantic guidance, attention-style modulation
SFG-SwinSR (Hossain et al., 10 May 2026)	Remote sensing single-image super-resolution	Blur-based low/high decomposition, spatial refinement, bottleneck gate
DSFC-Net (Zhang et al., 1 Feb 2026)	Rural road extraction	Dual encoder, Laplacian Pyramid-style high/low decomposition, channel recalibration
WaveSFNet (Cai et al., 24 Mar 2026)	Spatiotemporal prediction	Wavelet codec, spatial-frequency dual-domain translator, gated channel interaction
FGINet (Zhou et al., 30 Apr 2026)	AI-generated image detection	Band-masked frequency encoder, layer-wise gated frequency injection

In this broader sense, “frequency-spatial synergistic gated” refers to three recurring commitments: explicit or frequency-inspired decomposition, preservation of spatial structure, and adaptive fusion. The literature differs primarily in how “frequency” is parameterized and how the adaptive coupling is implemented.

2. Core architectural pattern

A recurrent structural motif is to estimate a low-frequency component, isolate a high-frequency residual, refine the residual spatially, and re-inject it through a learned gate. In SFG-SwinSR, the transformer tokens are reshaped into a 2D map, low-frequency content is estimated by a depthwise $5\times5$ blur, and the residual is formed explicitly as

$$\mathbf{F}_{LF} = \mathrm{DWConv}_{k \times k}(\mathbf{F}), \qquad \mathbf{F}_{HF} = \mathbf{F} - \mathbf{F}_{LF},$$

followed by spatial refinement and adaptive reinjection,

$$\mathbf{F}_{out} = \mathbf{F} + \mathbf{G} \odot \tilde{\mathbf{F}}_{HF}.$$

Here the “frequency” mechanism is spatial-domain low-pass/high-pass approximation rather than Fourier or wavelet analysis (Hossain et al., 10 May 2026).

A second recurrent motif is dual-path processing. In WaveSFNet, the latent translator computes a spatial branch

$$\mathbf{S} = \operatorname{DWConv}_{9\times 9}\left(\widetilde{\mathbf{Z}}^{\mathrm{p}}\right)$$

and a frequency branch

$$\mathbf{P} = \mathcal{F}^{-1}\!\left( \mathcal{F}\!\left(\widetilde{\mathbf{Z}}^{\mathrm{p}}\right)\odot \boldsymbol{\Psi} \right),$$

then fuses them by

$\mathbf{U} = \mathbf{S} + \mathbf{P},$

before a gated channel interaction stage,

$$\mathbf{G}= \operatorname{Conv}_{1\times1}\!\Big( \operatorname{SiLU}(\mathbf{U}^{(u)})\odot \operatorname{DWConv}_{3\times3}(\mathbf{U}^{(v)}) \Big).$$

This makes the gating act after dual-domain context extraction, not merely as a scalar branch selector (Cai et al., 24 Mar 2026).

A third motif is pixel-wise scale selection driven by local spectral proxies. In PFGNet, local cues from gradient magnitude, Laplacian magnitude, and local variance are mapped to per-pixel softmax gates,

$$\alpha_k(h,w) = \frac{\exp(Z_{g,k}(h,w))}{\sum_{j=1}^{K}\exp(Z_{g,j}(h,w))},$$

which weight center-suppressed large-kernel spatial responses,

$$\mathbf{Y}_k = \mathbf{P}_k - \tanh(\beta_k)\cdot (c * \mathbf{X}), \qquad \mathrm{PFG}(\mathbf{X}) = \sum_{k\in\mathcal{K}} \alpha_k \odot \mathbf{Y}_k.$$

This design treats gating as spatially adaptive receptive-field control conditioned on local frequency evidence (Cai et al., 24 Feb 2026).

Taken together, these examples show that FSG-Net-like systems are defined less by a single topology than by a recurrent computational grammar: frequency estimation, spatial preservation, and adaptive recombination.

3. Representative architectural realizations

In transformer super-resolution, the design is localized to the feed-forward layer rather than the attention layer. SFG-SwinSR preserves the Swin2SR backbone, including window-based self-attention and shifted windows, but replaces every standard FFN with Spatial-Frequency Gated Feed-Forward Network (SFG-FFN). The reported configuration uses shallow convolutional embedding, six Swin Transformer stages, six transformer layers per stage, 36 blocks total, window size 8, embedding dimension $C=180$, 6 attention heads per block, expansion ratio $$\mathbf{F}_{LF} = \mathrm{DWConv}_{k \times k}(\mathbf{F}), \qquad \mathbf{F}_{HF} = \mathbf{F} - \mathbf{F}_{LF},$$0, blur kernel size $$\mathbf{F}_{LF} = \mathrm{DWConv}_{k \times k}(\mathbf{F}), \qquad \mathbf{F}_{HF} = \mathbf{F} - \mathbf{F}_{LF},$$1, and gate reduction ratio $$\mathbf{F}_{LF} = \mathrm{DWConv}_{k \times k}(\mathbf{F}), \qquad \mathbf{F}_{HF} = \mathbf{F} - \mathbf{F}_{LF},$$2 (Hossain et al., 10 May 2026).

In segmentation, the pattern often appears as a dual-encoder system. DSFC-Net uses a ConvNeXt-v2 CNN branch for local boundaries and a Spatial-Frequency Hybrid Transformer (SFT) branch for global topology. Its Cross-Frequency Interaction Attention (CFIA) explicitly decomposes

$$\mathbf{F}_{LF} = \mathrm{DWConv}_{k \times k}(\mathbf{F}), \qquad \mathbf{F}_{HF} = \mathbf{F} - \mathbf{F}_{LF},$$3

and uses the original feature as query while taking keys and values from the high- and low-frequency branches. Fusion between the CNN and transformer streams is handled by Channel Feature Fusion Module (CFFM), which produces branch-specific channel weights from concatenated global descriptors (Zhang et al., 1 Feb 2026).

In spatiotemporal prediction, the same idea is embedded in recurrent-free latent translators. WaveSFNet combines a wavelet-based encoder-decoder with a spatial-frequency dual-domain translator, while PFGNet avoids explicit transform-domain processing and instead derives local spectral descriptors that gate multi-scale large-kernel spatial responses. The former emphasizes preserved wavelet subbands during downsampling and reconstruction; the latter emphasizes dynamic center-surround band-pass behavior and separable large-kernel efficiency (Cai et al., 24 Mar 2026, Cai et al., 24 Feb 2026).

In detection and forensics, the “synergy” often takes the form of controlled injection rather than symmetric fusion. FSGNet for IRSTD adds MIAM in the encoder, MFM on skip connections, and GPM with Global Semantic Guidance Flows in the decoder, so that frequency-aware skip refinement is paired with semantically stabilized decoding (Zhang et al., 26 Mar 2026). FGINet goes further by encoding a band-masked frequency token and injecting it into a frozen DINOv3 ViT-L/14 class token via a learnable layer-wise scalar gate,

$$\mathbf{F}_{LF} = \mathrm{DWConv}_{k \times k}(\mathbf{F}), \qquad \mathbf{F}_{HF} = \mathbf{F} - \mathbf{F}_{LF},$$4

which is explicitly designed to alleviate cross-domain representation conflict (Zhou et al., 30 Apr 2026).

4. Frequency parameterization and gating taxonomy

A common misconception is that “frequency-aware” necessarily implies Fourier analysis. The literature is more heterogeneous. SFG-SwinSR uses a trainable depthwise blur and subtraction to approximate low- and high-frequency components entirely in the spatial domain (Hossain et al., 10 May 2026). DSFC-Net uses a Laplacian Pyramid-style decomposition implemented by pooling, upsampling, and subtraction rather than a spectral transform (Zhang et al., 1 Feb 2026). PFGNet uses Sobel, Laplacian, and local variance as local spectral proxies rather than explicit transform coefficients (Cai et al., 24 Feb 2026).

Transform-domain variants are equally prominent, but they differ in what they keep and how they fuse it. FSGNet for IRSTD applies FFT to intermediate skip features, separates real and imaginary components,

$$\mathbf{F}_{LF} = \mathrm{DWConv}_{k \times k}(\mathbf{F}), \qquad \mathbf{F}_{HF} = \mathbf{F} - \mathbf{F}_{LF},$$5

and reconstructs refined features through

$$\mathbf{F}_{LF} = \mathrm{DWConv}_{k \times k}(\mathbf{F}), \qquad \mathbf{F}_{HF} = \mathbf{F} - \mathbf{F}_{LF},$$6

using frequency-domain modulation to suppress clutter (Zhang et al., 26 Mar 2026). FGINet uses Haar DWT, discards $$\mathbf{F}_{LF} = \mathrm{DWConv}_{k \times k}(\mathbf{F}), \qquad \mathbf{F}_{HF} = \mathbf{F} - \mathbf{F}_{LF},$$7, concatenates $$\mathbf{F}_{LF} = \mathrm{DWConv}_{k \times k}(\mathbf{F}), \qquad \mathbf{F}_{HF} = \mathbf{F} - \mathbf{F}_{LF},$$8, $$\mathbf{F}_{LF} = \mathrm{DWConv}_{k \times k}(\mathbf{F}), \qquad \mathbf{F}_{HF} = \mathbf{F} - \mathbf{F}_{LF},$$9, and $$\mathbf{F}_{out} = \mathbf{F} + \mathbf{G} \odot \tilde{\mathbf{F}}_{HF}.$$0, and applies independent band-wise masking before convolutional encoding of a single frequency token (Zhou et al., 30 Apr 2026).

The meaning of “gated” also varies. In some models it is a scalar branch mixer, as in DB-FGA-Net, where a global gate

$$\mathbf{F}_{out} = \mathbf{F} + \mathbf{G} \odot \tilde{\mathbf{F}}_{HF}.$$1

balances co-attended spatial features and FFT-derived frequency attention (Shreya et al., 23 Oct 2025). In others it is channel-wise bottleneck recalibration, as in SFG-SwinSR (Hossain et al., 10 May 2026). In FGINet, it is a layer-wise scalar injection coefficient (Zhou et al., 30 Apr 2026). In PFGNet, it is a pixel-wise softmax over receptive-field scales (Cai et al., 24 Feb 2026). In dual-domain fusion blocks such as FSCN and remote-sensing SFFNet, the gating behavior is realized by cross-domain attention weights rather than an explicit sigmoid mask, which suggests that “gated” in this family includes both multiplicative gates and attention-based selective routing (Lu et al., 2023, Yang et al., 2024).

This diversity implies that FSG-Net is a mechanism class, not a single operator family.

5. Empirical evidence across application domains

The design pattern has been reported across restoration, segmentation, prediction, detection, and forensic classification.

Model	Domain	Representative reported result
SFG-SwinSR	Remote sensing super-resolution	SpaceNet: 45.19 dB PSNR, 0.9852 SSIM, 0.0031 MAE
DSFC-Net	Rural road extraction	DeepGlobe: 81.44% F1, 68.68% IoU
WaveSFNet	Spatiotemporal prediction	TaxiBJ: 0.2870 MSE, 0.9859 SSIM, 39.88 PSNR
FSGNet	Infrared small target detection	NUDT-SIRST: 93.78 IoU, 94.14 nIoU, 99.26 Pa, 4.89 Fa
FGINet	AI-generated image detection	GenImage: 96.7% mAcc; Synthbuster: 94.3% mAcc

The ablation literature is especially informative. In SFG-SwinSR, the proposed backbone is not automatically superior under plain $$\mathbf{F}_{out} = \mathbf{F} + \mathbf{G} \odot \tilde{\mathbf{F}}_{HF}.$$2 training: SFG-FFN backbone + L1 only gives 42.63 dB, 0.9785, 0.0049, whereas the baseline FFN backbone + L1 only gives 43.56 dB, 0.9780, 0.0039. The gain appears only when the architecture is paired with structured supervision such as SSIM, edge, and frequency losses, culminating in 45.19 dB, 0.9852, 0.0031 on SpaceNet. This supports the authors’ claim that the architecture and loss design are complementary (Hossain et al., 10 May 2026).

In WaveSFNet, the full dual-domain model outperforms Spatial-only, Frequency-only, w/o TDI Block, Conv codec, and MLP mixer variants. On TaxiBJ, the full model reaches 0.2870 MSE, compared with 0.2949 for Spatial-only, 0.3014 for Frequency-only, 0.2962 without TDI, 0.3198 with Conv codec, and 0.2997 with MLP mixer. This indicates that the wavelet codec, dual-domain interaction, and gated channel mixing each contribute materially (Cai et al., 24 Mar 2026).

In DSFC-Net, removing the Transformer branch reduces performance from 69.93% F1 / 53.77% IoU to 68.10% / 51.63%; removing the CNN branch yields 69.56% / 53.33%; removing CFFM gives 69.63% / 53.41%. The strongest system therefore requires both branches and adaptive fusion rather than either component alone (Zhang et al., 1 Feb 2026).

In FGINet, ungated layer-wise injection is actively harmful: LFI (all) gives 68.9 Acc / 69.9 AP on Chameleon, while full LGFI all layers reaches 92.5 Acc / 95.7 AP. The same table shows that simple late Add, Concat, or Cross-Attention fusion is inferior to progressive gated injection. This is one of the clearest demonstrations that controlled cross-domain coupling matters more than mere multimodal availability (Zhou et al., 30 Apr 2026).

6. Limits, misconceptions, and plausible directions

The first limit is terminological. The exact acronym FSGNet is already occupied by “Frequency-Aware and Semantic Guidance Network” in IRSTD, so the phrase “Frequency-Spatial Synergistic Gated Network” is not the canonical expansion of that specific model (Zhang et al., 26 Mar 2026). A plausible implication is that the phrase functions more reliably as a descriptive category than as a single proper noun.

The second limit is conceptual overreach. “Frequency-aware” is sometimes used for explicit FFT or wavelet processing, but in other cases it denotes blur-residual approximation, Laplacian-style decomposition, or local derivative-based proxies. Equating all such systems with explicit spectral learning would be inaccurate (Hossain et al., 10 May 2026, Zhang et al., 1 Feb 2026, Cai et al., 24 Feb 2026). Likewise, “gated” does not always mean a sigmoid mask; it may refer to scalar injection, channel recalibration, cross-attention weighting, or softmax scale selection (Zhou et al., 30 Apr 2026, Lu et al., 2023).

The third limit is evidential scope. Several results are domain-specific or benchmark-specific. SFG-SwinSR is evaluated on synthetic degradation for SpaceNet and a selected subset of SEN2VEN$$\mathbf{F}_{out} = \mathbf{F} + \mathbf{G} \odot \tilde{\mathbf{F}}_{HF}.$$3S, and broader real-world validation is explicitly noted as still needed (Hossain et al., 10 May 2026). DSFC-Net reports 58.06 M parameters and 122.57 G FLOPs, indicating that synergistic dual-encoder designs can be expensive (Zhang et al., 1 Feb 2026). FGINet reports strong accuracy but does not provide FLOPs, parameter counts, or latency analysis in the supplied description (Zhou et al., 30 Apr 2026). These constraints suggest that frequency-spatial gating is not a free architectural primitive; it trades off inductive bias, cost, and transfer.

A final misconception is to treat these systems as interchangeable. They are not. Some are feed-forward replacements inside a transformer block, some are dual-encoder fusion networks, some are recurrent-free latent translators, and some are guided injection mechanisms for frozen vision foundation models. What unifies them is not topology but the claim that local spatial structure and frequency-aware context are complementary and should be combined adaptively.

A plausible direction emerging from this literature is finer-grained control over where and how cross-domain information is injected: more explicit multi-level frequency hierarchies, spatially varying gates beyond global scalars, stronger real-world validation under domain shift, and tighter coupling between architectural priors and supervision. The existing record already shows that when frequency modeling, spatial preservation, and adaptive fusion are aligned, the resulting systems can outperform either spatial-only or frequency-only baselines across a wide range of tasks.