Frequency-Enhanced Network (FENet)

• FENet is a family of neural architectures that decompose, enhance, and reintegrate frequency bands to extract robust spatial and temporal features.
• It employs transforms like FFT, wavelet, and DCT to enable explicit spectral control and cross-band fusion across diverse applications.
• FENet designs have demonstrated state-of-the-art results in segmentation, restoration, medical diagnosis, and sequence modeling while ensuring computational efficiency.

Frequency-Enhanced Network (FENet) denotes a class of neural architectures in which representation learning explicitly manipulates frequency content—typically by separating low-, mid-, and high-frequency components, enhancing selected bands, and reintegrating them with spatial, temporal, or graph-structural features. Across recent literature, the term is not used uniformly: some papers formalize a model named FENet, whereas others present AFENet, FE-UNet, FEDIN, FEDSNet, or a descriptive “frequency-domain enhanced” framework that instantiates the same principle in segmentation, restoration, medical diagnosis, recommendation, and sequence modeling (Huo et al., 6 Feb 2025, Liu et al., 4 Sep 2025, Ye et al., 2021, Yan et al., 2024).

1. Terminology, scope, and naming

The literature uses “FENet” in two distinct ways. In a narrow sense, it is an explicit model name, as in “FENet: A Frequency Extraction Network for Obstructive Sleep Apnea Detection” and the self-supervised fMRI framework titled “Frequency-Enhanced Network (FENet)” (Ye et al., 2021, Liu et al., 4 Sep 2025). In a broader sense, it functions as a conceptual label for architectures that explicitly enhance or exploit frequency-domain information. This broader usage is explicit in the FE-UNet paper, which states that the exact term “FENet” is not formalized as a standalone model name there; instead, FE-UNet is presented as a concrete instantiation of a Frequency-Enhanced Network built by embedding WSPM and FE-RFB into a U-Net using SAM2/Hiera-L (Huo et al., 6 Feb 2025).

This naming heterogeneity recurs across other domains. The deraining and remote-sensing segmentation papers both use the formal name Adaptive Frequency Enhancement Network (AFENet), while simultaneously fitting the broader FENet idea because they explicitly decompose, modulate, and fuse frequency-domain information with spatial features (Yan et al., 2024, Gao et al., 3 Apr 2025). Likewise, the CTR model FEDIN and the few-shot classifier FEDSNet are not named “FENet,” but each is explicitly frequency-enhanced: FEDIN uses a target-aware frequency-domain branch, whereas FEDSNet uses DCT-based low-pass decomposition and dual subspaces for structural stabilization (Dai et al., 3 May 2026, Wang et al., 16 Apr 2026).

Accordingly, FENet is best treated as a research category rather than a single canonical architecture. Within that category, the common denominator is explicit spectral control: the model does not merely hope that a backbone implicitly learns useful frequency structure, but introduces dedicated mechanisms for decomposition, filtering, cross-band interaction, or spectral supervision.

2. Core methodological principles

A recurring design pattern is explicit band separation followed by band-specific enhancement and cross-domain fusion. In FE-UNet, this pattern is formalized through the Wavelet-Guided Spectral Pooling Module (WSPM) and the Frequency Domain Enhanced Receptive Field Block (FE-RFB). The paper motivates these modules by reporting that CNNs are weak on low-frequency signals and relatively stronger on mid-to-high frequencies, whereas the human visual system exhibits band-pass behavior with peak sensitivity in mid frequencies. WSPM therefore enhances low-frequency structure via cascaded deep wavelet convolution and rebalances spectra through spectral pooling, with the mixed response written as $$\tilde{V} = \lambda f_{lp}(Z) + (1-\lambda) f_{hp}(Z)$$; in implementation, two parallel SPF branches use $\lambda = 0.7$ and $0.8$ to emphasize low-mid content (Huo et al., 6 Feb 2025).

In graph-based psychiatric disorder detection, FENet follows the same logic in a different mathematical setting. The model constructs time-domain and frequency-domain graph views over the same brain topology, applies Graph Fourier Transform using $$\mathbf{L}=\mathbf{U}\mathbf{\Lambda}\mathbf{U}^\top$$, selects low-, mid-, and high-frequency graph bands through eigenvalue thresholds at the 20th and 80th percentiles, and then fuses embeddings as $$\mathbf{Z}_{\mathrm{TF}}=\frac{1}{2}(\mathbf{Z}_{\mathrm{T}}+\mathbf{Z}_{\mathrm{F}})$$ under a CCA-inspired non-contrastive objective (Liu et al., 4 Sep 2025). In this setting, “frequency enhancement” means emphasizing disease-relevant graph frequencies while preserving biological topology.

Other instantiations differ mainly in the transform used and the point at which spectral processing enters the network. AFENet for deraining uses stride-based convolutional decomposition into low-, mid-, and high-frequency branches rather than a fixed DCT or wavelet split; DFENet for demosaicking uses FFT-domain selectors to send different spectral regions either to spatial synthesis or CFA-guided suppression; FEMSN uses learnable complex FFT masks in FADEL and again inside MSTFF blocks; and the physical-guided recurrent model applies learned complex Fourier filters inside residual Fourier modules and combines them with an $H^1$ loss that up-weights high-frequency errors (Yan et al., 2024, Liu et al., 20 Mar 2025, Yuan et al., 7 May 2025, Zhao et al., 2024). FEDSNet uses a different frequency logic again: a 2D DCT and a low-pass mask isolate global structural information before truncated SVD constructs a frequency-domain subspace (Wang et al., 16 Apr 2026).

Not all FENet-style models require explicit FFT or DCT layers. The OSA-detection FENet uses learnable multi-frequency dilated convolutions, parameterized by $freq(k_j)=freq_s/d_j$, to form a frequency extractor over RR-interval sequences and then maps one observed epoch to multiple output epochs (Ye et al., 2021). The fatigue-life predictor uses FNet blocks, where parameterless Fourier mixing replaces self-attention in one of several parallel branches (Chen et al., 2024). A plausible implication is that “frequency enhancement” in current usage refers less to any single transform family than to a design commitment: frequency information must be structurally exposed and operationalized.

3. Representative architectural instantiations

Model	Domain	Frequency mechanism
FE-UNet (Huo et al., 6 Feb 2025)	Image segmentation	Haar wavelets, spectral pooling, FE-RFB
AFENet (Yan et al., 2024)	Single-image deraining	Stride-based low/mid/high decomposition, transformer FEM, PA-guided interaction
FENet (Liu et al., 4 Sep 2025)	fMRI psychiatric detection	GFT, band filters, TGNN/FGNN, CCA-guided SSL
FENet (Ye et al., 2021)	OSA detection from RR intervals	Multi-frequency dilated CNN, one-to-multiple decoding
DFENet (Liu et al., 20 Mar 2025)	Image demosaicking	FFT selectors, CFA-guided false-frequency suppression
FEMSN (Yuan et al., 7 May 2025)	Fault diagnosis	FFT-based FADEL, MSTFF time-frequency fusion
AFENet (Gao et al., 3 Apr 2025)	Remote-sensing segmentation	Adaptive FFT window masks, cross-attention, selective fusion
FEDSNet (Wang et al., 16 Apr 2026)	Few-shot fine-grained classification	DCT low-pass branch, dual SVD subspaces, adaptive gating
FEDIN (Dai et al., 3 May 2026)	CTR prediction	Target-aware spectral masking, complex MLP filtering
SFD-Mamba2Net (Mu et al., 10 Sep 2025)	Coronary segmentation	Multi-level Haar DWT/IWT PHFP in decoder

These models occupy different points in the architectural design space. FE-UNet, AFENet for remote sensing, and SFD-Mamba2Net place frequency modules inside encoder-decoder segmentation systems, where spectral enhancement is tied directly to multi-scale feature fusion and boundary recovery. AFENet for deraining and DFENet for demosaicking instead treat frequency processing as a restoration mechanism: they use per-band enhancement, cross-band guidance, or explicit false-frequency suppression to recover clean images from corrupted observations.

The graph-SSL FENet, the OSA FENet, and FEDIN show that the idea transfers beyond images. In fMRI, frequency enhancement is applied to graph spectra over ROIs; in RR-interval OSA screening, it is implemented through multi-dilation temporal filters and a one-to-multiple decoder; in CTR prediction, it becomes target-aware spectrum filtering over behavior sequences. The fatigue-life predictor and the physical-guided recurrent network extend the same principle to scientific and engineering time series, combining Fourier mixing with recurrent or multi-view structure (Zhao et al., 2024, Chen et al., 2024).

4. Empirical performance across domains

In segmentation, FE-UNet reports state-of-the-art performance on several tasks. On MAS3K it achieves mIoU $=0.815$, $S_\alpha=0.900$, $F^w_\beta=0.848$, $\lambda = 0.7$0, and MAE $\lambda = 0.7$1; on RUWI it reaches mIoU $\lambda = 0.7$2 and $\lambda = 0.7$3; on Kvasir-SEG it reports mDice $\lambda = 0.7$4 and mIoU $\lambda = 0.7$5 (Huo et al., 6 Feb 2025). In high-resolution remote-sensing segmentation, AFENet reports mIoU $\lambda = 0.7$6 and mF1 $\lambda = 0.7$7 on ISPRS Vaihingen, mIoU $\lambda = 0.7$8 and mF1 $\lambda = 0.7$9 on ISPRS Potsdam, and mIoU $0.8$0 on LoveDA (Gao et al., 3 Apr 2025). In ICA coronary segmentation, SFD-Mamba2Net reports Dice $0.8$1, IoU $0.8$2, HD95 $0.8$3 px, and ASSD $0.8$4 px, together with stenosis-detection results of TPR $0.8$5, PPV $0.8$6, ARMSE $0.8$7, and RRMSE $0.8$8 (Mu et al., 10 Sep 2025).

In restoration, AFENet for deraining reports on Rain13K tests: Test100 $0.8$9, Rain100H $$\mathbf{L}=\mathbf{U}\mathbf{\Lambda}\mathbf{U}^\top$$0, Rain100L $$\mathbf{L}=\mathbf{U}\mathbf{\Lambda}\mathbf{U}^\top$$1, Test1200 $$\mathbf{L}=\mathbf{U}\mathbf{\Lambda}\mathbf{U}^\top$$2, Test2800 $$\mathbf{L}=\mathbf{U}\mathbf{\Lambda}\mathbf{U}^\top$$3, and average $$\mathbf{L}=\mathbf{U}\mathbf{\Lambda}\mathbf{U}^\top$$4 PSNR / $$\mathbf{L}=\mathbf{U}\mathbf{\Lambda}\mathbf{U}^\top$$5 SSIM; on real scenes it reports NIQE/BRISQUE $$\mathbf{L}=\mathbf{U}\mathbf{\Lambda}\mathbf{U}^\top$$6 on Real15 and $$\mathbf{L}=\mathbf{U}\mathbf{\Lambda}\mathbf{U}^\top$$7 on Real300 (Yan et al., 2024). DFENet for demosaicking reports, at $$\mathbf{L}=\mathbf{U}\mathbf{\Lambda}\mathbf{U}^\top$$8, Kodak $$\mathbf{L}=\mathbf{U}\mathbf{\Lambda}\mathbf{U}^\top$$9, Set14 $$\mathbf{Z}_{\mathrm{TF}}=\frac{1}{2}(\mathbf{Z}_{\mathrm{T}}+\mathbf{Z}_{\mathrm{F}})$$0, Urban100 $$\mathbf{Z}_{\mathrm{TF}}=\frac{1}{2}(\mathbf{Z}_{\mathrm{T}}+\mathbf{Z}_{\mathrm{F}})$$1, and MIT moiré $$\mathbf{Z}_{\mathrm{TF}}=\frac{1}{2}(\mathbf{Z}_{\mathrm{T}}+\mathbf{Z}_{\mathrm{F}})$$2, while LineSet37 is introduced specifically to test recovery of challenging line patterns and color moiré (Liu et al., 20 Mar 2025).

In biomedical and sequence settings, the graph-SSL FENet reports on ABIDE ACC $$\mathbf{Z}_{\mathrm{TF}}=\frac{1}{2}(\mathbf{Z}_{\mathrm{T}}+\mathbf{Z}_{\mathrm{F}})$$3, AUC $$\mathbf{Z}_{\mathrm{TF}}=\frac{1}{2}(\mathbf{Z}_{\mathrm{T}}+\mathbf{Z}_{\mathrm{F}})$$4, Recall $$\mathbf{Z}_{\mathrm{TF}}=\frac{1}{2}(\mathbf{Z}_{\mathrm{T}}+\mathbf{Z}_{\mathrm{F}})$$5, and F1 $$\mathbf{Z}_{\mathrm{TF}}=\frac{1}{2}(\mathbf{Z}_{\mathrm{T}}+\mathbf{Z}_{\mathrm{F}})$$6; on ADHD-200 it reports ACC $$\mathbf{Z}_{\mathrm{TF}}=\frac{1}{2}(\mathbf{Z}_{\mathrm{T}}+\mathbf{Z}_{\mathrm{F}})$$7, AUC $$\mathbf{Z}_{\mathrm{TF}}=\frac{1}{2}(\mathbf{Z}_{\mathrm{T}}+\mathbf{Z}_{\mathrm{F}})$$8, Recall $$\mathbf{Z}_{\mathrm{TF}}=\frac{1}{2}(\mathbf{Z}_{\mathrm{T}}+\mathbf{Z}_{\mathrm{F}})$$9, and F1 $H^1$0, with accuracy gains of $H^1$1 on ABIDE and $H^1$2 on ADHD over CCA-SSG in the 20% labeled regime (Liu et al., 4 Sep 2025). The RR-interval OSA FENet reports, in continuous detection on PAD*, Acc $H^1$3, Rec $H^1$4, Pre $H^1$5, and Spe $H^1$6; in discontinuous detection with duty cycle $H^1$7 on PAD-UCDSAD, it reports Acc $H^1$8, Rec $H^1$9, Pre $freq(k_j)=freq_s/d_j$0, and Spe $freq(k_j)=freq_s/d_j$1 (Ye et al., 2021). FEMSN reports, on PU, mean accuracy $freq(k_j)=freq_s/d_j$2 at SNR $freq(k_j)=freq_s/d_j$3 dB and $freq(k_j)=freq_s/d_j$4 at $freq(k_j)=freq_s/d_j$5 dB, and on SEU, $freq(k_j)=freq_s/d_j$6 at $freq(k_j)=freq_s/d_j$7 dB and $freq(k_j)=freq_s/d_j$8 at $freq(k_j)=freq_s/d_j$9 dB and $=0.815$0 dB (Yuan et al., 7 May 2025).

In recommendation and few-shot recognition, FEDIN reports GAUC/AUC of $=0.815$1 on Tmall, $=0.815$2 on Alipay, and $=0.815$3 on Taobao, with statistically significant gains over the best baselines (Dai et al., 3 May 2026). FEDSNet reports on Stanford Cars, with a ResNet-12 backbone, $=0.815$4 in 1-shot and $=0.815$5 in 5-shot, and on CUB $=0.815$6 and $=0.815$7 respectively (Wang et al., 16 Apr 2026). The physical-guided, frequency-enhanced recurrent model reports $=0.815$8 MAE and $=0.815$9 SSIM on TaxiBJ, $S_\alpha=0.900$0 CSI-M on SEVIR, and $S_\alpha=0.900$1 N-MSE on Navier–Stokes, while using approximately $S_\alpha=0.900$2–$S_\alpha=0.900$3M parameters (Zhao et al., 2024).

5. Training regimes, efficiency, and reproducibility

Despite their diversity, FENet-style models often pursue spectral enhancement under explicit efficiency constraints. FE-UNet freezes Hiera-L from SAM2, inserts a lightweight adapter, reduces each level to 64 channels, and trains in PyTorch with AdamW, initial learning rate $S_\alpha=0.900$4, cosine decay, batch size $S_\alpha=0.900$5, and 20 epochs on $S_\alpha=0.900$6 inputs; however, the paper does not report parameters, FLOPs, or a public repository link (Huo et al., 6 Feb 2025). The graph-SSL FENet emphasizes computational efficiency more directly: its FGNN replaces heavier spectral GCN operations with FGO layers and reports complexity $S_\alpha=0.900$7, with AdamW at learning rate $S_\alpha=0.900$8 and five repeated runs on a Tesla P100 16 GB GPU (Liu et al., 4 Sep 2025).

Several papers tie frequency enhancement to resource efficiency in deployment. The OSA FENet is explicitly designed for energy-constrained wearables and models sensor consumption as $S_\alpha=0.900$9 versus $F^w_\beta=0.848$0, thereby reducing sensor operation to one-third while preserving continuous output through one-to-multiple decoding (Ye et al., 2021). The fatigue-life predictor uses FNet because Fourier mixing scales as $F^w_\beta=0.848$1 rather than $F^w_\beta=0.848$2 self-attention over sequence length (Chen et al., 2024). SFD-Mamba2Net likewise stresses that PHFP uses fixed Haar filters and lightweight depthwise $F^w_\beta=0.848$3 convolutions, while CASE is parameter-free and AA-DS Mamba2 leverages linear-time chunked parallelism under State Space Duality (Mu et al., 10 Sep 2025).

Other implementations make the efficiency profile explicit through parameter and runtime reporting. DFENet reports approximately $F^w_\beta=0.848$4M parameters for the full model, $F^w_\beta=0.848$5M for DFENet-S, and $F^w_\beta=0.848$6M for DFENet-T, while using FFT/IFFT with lightweight $F^w_\beta=0.848$7 convolutions in the suppression path and a stagewise training scheme with $F^w_\beta=0.848$8 over 500,000 iterations (Liu et al., 20 Mar 2025). FEDSNet reports, for 5-way 5-shot on a single RTX 3090, that with a ResNet-12 backbone it uses $F^w_\beta=0.848$9M parameters, $\lambda = 0.7$00G FLOPs, $\lambda = 0.7$01 MB peak memory, and $\lambda = 0.7$02 ms per task; with Conv-4, it uses $\lambda = 0.7$03M parameters and $\lambda = 0.7$04G FLOPs (Wang et al., 16 Apr 2026). AFENet for remote sensing reports $\lambda = 0.7$05M parameters and $\lambda = 0.7$06G FLOPs for a $\lambda = 0.7$07 input, and releases code (Gao et al., 3 Apr 2025). FEDIN also releases code and specifies FuxiCTR, Adam, learning rate $\lambda = 0.7$08, batch size $\lambda = 0.7$09, embedding dimension $\lambda = 0.7$10, and maximum sequence length $\lambda = 0.7$11 (Dai et al., 3 May 2026).

Reproducibility remains uneven. Code is explicitly available for DFENet, AFENet for remote sensing, FEDIN, and SFD-Mamba2Net, whereas FE-UNet and the fMRI FENet do not release public repositories in the paper text (Liu et al., 20 Mar 2025, Gao et al., 3 Apr 2025, Dai et al., 3 May 2026, Mu et al., 10 Sep 2025).

6. Limitations, misconceptions, and future directions

A common misconception is that a FENet must be an FFT-based image model. The literature does not support that restriction. Some models use FFT or DFT masks; others use wavelets, DCT, graph Fourier transforms, FNet-style Fourier mixing, or multi-dilated convolutions as an implicit frequency extractor (Huo et al., 6 Feb 2025, Liu et al., 4 Sep 2025, Ye et al., 2021, Wang et al., 16 Apr 2026). A second misconception is that “segment-anything capability” implies prompt-interactive segmentation. In FE-UNet, that capability comes from reusing SAM2’s Hiera backbone for representation power; the paper explicitly states that FE-UNet is not prompt-interactive and that class-specific performance depends on the fine-tuning dataset (Huo et al., 6 Feb 2025).

The main technical limitations are also recurrent. FE-UNet, AFENet for deraining, and AFENet for remote sensing all state that frequency hyperparameters matter: poor settings can over-suppress high-frequency detail, under-enhance low-frequency structure, or fail to adapt across scene types (Huo et al., 6 Feb 2025, Yan et al., 2024, Gao et al., 3 Apr 2025). FEDSNet notes that a fixed low-pass cutoff can discard subtle discriminative high-frequency structure, especially in fine-grained categories dominated by minute local parts (Wang et al., 16 Apr 2026). The ICA paper notes limited dataset scale and the use of single-view static frames rather than multi-view or time-resolved clinical data (Mu et al., 10 Sep 2025). The fatigue-life predictor reports strong extrapolation behavior but does not include ablations isolating the contribution of its FNet branch or direct quantitative comparisons to branch-removed variants (Chen et al., 2024).

Several future directions appear repeatedly. The fMRI FENet highlights harmonization and domain adaptation for cross-site variability, as well as stronger interpretability in the spectral pathway (Liu et al., 4 Sep 2025). The wearable OSA FENet points toward adaptive sensing and possibly dynamic duty cycles (Ye et al., 2021). AFENet for remote sensing explicitly identifies multimodal extension, such as optical plus SAR, and lower-cost architectures for real-time UAV use (Gao et al., 3 Apr 2025). FEDIN suggests that target-aware frequency conditioning may generalize beyond CTR to other sequential settings where contextual relevance changes the useful spectrum (Dai et al., 3 May 2026). A plausible implication is that the field is moving from fixed band decomposition toward input-adaptive and task-conditioned spectral operators, while simultaneously seeking tighter biological, physical, or geometric priors to prevent frequency enhancement from becoming a generic denoising heuristic rather than a principled inductive bias.