---
title: Frequency-Enhanced Network (FENet)
url: https://www.emergentmind.com/topics/frequency-enhanced-network-fenet
type: topic
---

# Frequency-Enhanced Network (FENet)

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 [2502.03829][2509.10524][2101.02873][2407.14292].

## 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)” [2101.02873][2509.10524]. 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 [2502.03829].

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 [2407.14292][2504.02647]. 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 [2605.01726][2604.14958].

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 [2502.03829].

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 [2509.10524]. 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 [2407.14292][2503.15800][2505.06285][2405.14504]. 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 [2604.14958].

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 [2101.02873]. The fatigue-life predictor uses FNet blocks, where parameterless Fourier mixing replaces self-attention in one of several parallel branches [2405.07507]. 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 [2502.03829] | Image segmentation | Haar wavelets, spectral pooling, FE-RFB |
| AFENet [2407.14292] | Single-image deraining | Stride-based low/mid/high decomposition, transformer FEM, PA-guided interaction |
| FENet [2509.10524] | fMRI psychiatric detection | GFT, band filters, TGNN/FGNN, CCA-guided SSL |
| FENet [2101.02873] | OSA detection from RR intervals | Multi-frequency dilated CNN, one-to-multiple decoding |
| DFENet [2503.15800] | Image demosaicking | FFT selectors, CFA-guided false-frequency suppression |
| FEMSN [2505.06285] | Fault diagnosis | FFT-based FADEL, MSTFF time-frequency fusion |
| AFENet [2504.02647] | Remote-sensing segmentation | Adaptive FFT window masks, cross-attention, selective fusion |
| FEDSNet [2604.14958] | Few-shot fine-grained classification | DCT low-pass branch, dual SVD subspaces, adaptive gating |
| FEDIN [2605.01726] | CTR prediction | Target-aware spectral masking, complex MLP filtering |
| SFD-Mamba2Net [2509.08934] | 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 [2405.14504][2405.07507].

## 4. Empirical performance across domains

In segmentation, FE-UNet reports state-of-the-art performance on several tasks. On MAS3K it achieves mIoU $=0.815$, $S_\alpha=0.900$, $F^w_\beta=0.848$, $mE_\phi=0.928$, and MAE $=0.022$; on RUWI it reaches mIoU $=0.914$ and $F^w_\beta=0.936$; on Kvasir-SEG it reports mDice $=0.929$ and mIoU $=0.883$ [2502.03829]. In high-resolution remote-sensing segmentation, AFENet reports mIoU $84.55$ and mF1 $91.54$ on ISPRS Vaihingen, mIoU $87.50$ and mF1 $93.24$ on ISPRS Potsdam, and mIoU $54.83$ on LoveDA [2504.02647]. In ICA coronary segmentation, SFD-Mamba2Net reports Dice $88.10\%$, IoU $80.81\%$, HD95 $7.57$ px, and ASSD $1.25$ px, together with stenosis-detection results of TPR $=0.60$, PPV $=0.64$, ARMSE $=0.14$, and RRMSE $=0.30$ [2509.08934].

In restoration, AFENet for deraining reports on Rain13K tests: Test100 $30.50/0.918$, Rain100H $31.22/0.901$, Rain100L $37.66/0.978$, Test1200 $33.13/0.925$, Test2800 $33.82/0.944$, and average $33.27$ PSNR / $0.933$ SSIM; on real scenes it reports NIQE/BRISQUE $15.90/23.18$ on Real15 and $15.01/26.34$ on Real300 [2407.14292]. DFENet for demosaicking reports, at $\sigma=0$, Kodak $43.39/0.9905/0.9899/0.0027$, Set14 $44.74/0.9893/0.9875/0.0024$, Urban100 $41.25/0.9871/0.9832/0.0025$, and MIT moiré $37.82/0.9461/0.9362/0.0230$, while LineSet37 is introduced specifically to test recovery of challenging line patterns and color moiré [2503.15800].

In biomedical and sequence settings, the graph-SSL FENet reports on ABIDE ACC $62.5\pm3.9$, AUC $64.7\pm4.1$, Recall $63.3\pm4.8$, and F1 $69.9\pm4.8$; on ADHD-200 it reports ACC $67.8\pm4.1$, AUC $69.7\pm4.4$, Recall $66.7\pm2.9$, and F1 $67.0\pm5.2$, with accuracy gains of $+6.5\%$ on ABIDE and $+11.2\%$ on ADHD over CCA-SSG in the 20% labeled regime [2509.10524]. The RR-interval OSA FENet reports, in continuous detection on PAD*, Acc $0.9922$, Rec $0.9925$, Pre $0.9985$, and Spe $0.9902$; in discontinuous detection with duty cycle $1/3$ on PAD-UCDSAD, it reports Acc $0.9567$, Rec $0.9608$, Pre $0.9793$, and Spe $0.9457$ [2101.02873]. FEMSN reports, on PU, mean accuracy $72.99\%$ at SNR $-10$ dB and $99.11\%$ at $-2$ dB, and on SEU, $95.51\%$ at $-10$ dB and $100\%$ at $-4$ dB and $-2$ dB [2505.06285].

In recommendation and few-shot recognition, FEDIN reports GAUC/AUC of $0.9658/0.9666$ on Tmall, $0.9335/0.9320$ on Alipay, and $0.9740/0.9729$ on Taobao, with statistically significant gains over the best baselines [2605.01726]. FEDSNet reports on Stanford Cars, with a ResNet-12 backbone, $85.04\pm0.19$ in 1-shot and $95.75\pm0.07$ in 5-shot, and on CUB $80.23\pm0.20$ and $90.78\pm0.11$ respectively [2604.14958]. The physical-guided, frequency-enhanced recurrent model reports $14.88$ MAE and $0.9851$ SSIM on TaxiBJ, $47.02$ CSI-M on SEVIR, and $14.89\times10^{-2}$ N-MSE on Navier–Stokes, while using approximately $3.4$–$3.8$M parameters [2405.14504].

## 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 $0.001$, cosine decay, batch size $12$, and 20 epochs on $350\times350$ inputs; however, the paper does not report parameters, FLOPs, or a public repository link [2502.03829]. The graph-SSL FENet emphasizes computational efficiency more directly: its FGNN replaces heavier spectral GCN operations with FGO layers and reports complexity $O(NK\log N + 3NK^2)$, with AdamW at learning rate $10^{-5}$ and five repeated runs on a Tesla P100 16 GB GPU [2509.10524].

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 $C_{STD}=(I_{PPG}+I_{BG})T$ versus $C_{FENet}=(\frac{1}{3}I_{PPG}+I_{BG})T$, thereby reducing sensor operation to one-third while preserving continuous output through one-to-multiple decoding [2101.02873]. The fatigue-life predictor uses FNet because Fourier mixing scales as $O(n\log n)$ rather than $O(n^2)$ self-attention over sequence length [2405.07507]. SFD-Mamba2Net likewise stresses that PHFP uses fixed Haar filters and lightweight depthwise $5\times5$ convolutions, while CASE is parameter-free and AA-DS Mamba2 leverages linear-time chunked parallelism under State Space Duality [2509.08934].

Other implementations make the efficiency profile explicit through parameter and runtime reporting. DFENet reports approximately $11.58$M parameters for the full model, $3.37$M for DFENet-S, and $0.12$M for DFENet-T, while using FFT/IFFT with lightweight $1\times1$ convolutions in the suppression path and a stagewise training scheme with $L=L_{rec}+0.01L_{fft}$ over 500,000 iterations [2503.15800]. FEDSNet reports, for 5-way 5-shot on a single RTX 3090, that with a ResNet-12 backbone it uses $13.30$M parameters, $352.36$G FLOPs, $697.39$ MB peak memory, and $147.28$ ms per task; with Conv-4, it uses $5.24$M parameters and $10.34$G FLOPs [2604.14958]. AFENet for remote sensing reports $20.23$M parameters and $25.59$G FLOPs for a $3\times512\times512$ input, and releases code [2504.02647]. FEDIN also releases code and specifies FuxiCTR, Adam, learning rate $5\times10^{-4}$, batch size $2048$, embedding dimension $32$, and maximum sequence length $100$ [2605.01726].

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 [2503.15800][2504.02647][2605.01726][2509.08934].

## 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 [2502.03829][2509.10524][2101.02873][2604.14958]. 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 [2502.03829].

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 [2502.03829][2407.14292][2504.02647]. 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 [2604.14958]. The ICA paper notes limited dataset scale and the use of single-view static frames rather than multi-view or time-resolved clinical data [2509.08934]. 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 [2405.07507].

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 [2509.10524]. The wearable OSA FENet points toward adaptive sensing and possibly dynamic duty cycles [2101.02873]. AFENet for remote sensing explicitly identifies multimodal extension, such as optical plus SAR, and lower-cost architectures for real-time UAV use [2504.02647]. FEDIN suggests that target-aware frequency conditioning may generalize beyond CTR to other sequential settings where contextual relevance changes the useful spectrum [2605.01726]. 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.

Source: https://www.emergentmind.com/topics/frequency-enhanced-network-fenet