Frequency-Adaptive Domain Fusion (FADF)

Updated 16 January 2026

FADF is a frequency-adaptive fusion strategy that decomposes data into low- and high-frequency components using adaptive transforms.
It leverages learnable wavelet, FFT, and DCT filters combined with gating and attention mechanisms to merge features from different domains.
The approach achieves state-of-the-art results in tasks such as image segmentation, deblurring, multimodal fusion, and medical imaging by improving robustness and detail preservation.

Frequency-Adaptive Domain Fusion (FADF) refers to a class of architectures and modules that perform data-driven, task-specific fusion of features, signals, or representations originating from distinct domains (e.g., spatial or frequency domains, multiple modalities, different datasets, or source–target pairs). FADF strategies utilize adaptive mechanisms (learnable wavelet/FFT/DCT transforms, gating networks, similarity or attention mappings) to separate, compare, and intelligently combine low- and high-frequency information for improved generalization, robustness, and detail preservation. FADF frameworks have achieved state-of-the-art results in image segmentation, image restoration, multimodal fusion, medical image processing, pan-sharpening, stereo matching, time series analysis, and more.

1. Foundational Principles of FADF

At the core of FADF is the notion that frequency content—especially the separation of low-frequency (domain-invariant) and high-frequency (domain-sensitive) components—encodes critical information for robust cross-domain generalization and fine-detail preservation. FADF modules typically:

Decompose input features or images into frequency subbands via wavelet, FFT, DCT, or learned filters (e.g., $\mathcal{W}_{\rm low}, \mathcal{W}_{\rm high}$ , AdaWAT, FDGM) (Wang et al., 9 Jan 2026, Gao et al., 20 Feb 2025, Wang et al., 21 Aug 2025, He et al., 2024).
Compute domain or modality-specific prototypes or statistics (e.g., global average pooling of wavelet bands) (Wang et al., 9 Jan 2026).
Establish similarity, attention, or gating metrics that quantify the affinity between the test/target domain's frequency signature and those of learned source domains or modalities (Wang et al., 9 Jan 2026, Xu et al., 4 Dec 2025).
Softly weight, select, or dynamically blend features, experts, or predictions associated with different domains, using temperature-scaled softmax, mixture-of-experts gates, spatial-frequency masks, or low-rank attention (Wang et al., 9 Jan 2026, He et al., 2024, Xu et al., 4 Dec 2025).
Ensure that fusion is performed adaptively—often at test time—so the network responds to the statistics of the unseen data or task (Wang et al., 9 Jan 2026, Li et al., 5 Dec 2025).

FADF lends itself to domain generalization, multimodal fusion, and adaptive enhancement, capitalizing on spectral information that is less susceptible to certain forms of domain shift (e.g., illumination, color cast, sensor variations) while allowing for fine, context-sensitive interpolation across domains.

2. Mathematical Formulations and Algorithmic Implementations

FADF systems span a broad range of mathematical instantiations, often integrating classic signal transforms with deep neural mechanisms. Selected formalizations include:

Precompute spectral prototypes per domain $k$ :

$\bar{F}_{\rm low}^k = \frac{1}{N_k}\sum_{i=1}^{N_k} \mathrm{GAP}\left(\mathcal{W}_{\rm low}(F_i^k)\right), \quad \bar{F}_{\rm high}^k = \frac{1}{N_k}\sum_{i=1}^{N_k} \mathrm{GAP}\left(\mathcal{W}_{\rm high}(F_i^k)\right)$

For a test sample, compute cosine similarity to domain prototypes in both bands:

$s_k = \frac{1}{2}\left[ \frac{\bar{F}_{\rm low}^{\rm test}\cdot\bar{F}_{\rm low}^k}{\|\bar{F}_{\rm low}^{\rm test}\|\|\bar{F}_{\rm low}^k\|} + \frac{\bar{F}_{\rm high}^{\rm test}\cdot\bar{F}_{\rm high}^k}{\|\bar{F}_{\rm high}^{\rm test}\|\|\bar{F}_{\rm high}^k\|} \right]$

Assign fusion weights by softmax:

$w_k = \frac{\exp(s_k/\tau)}{\sum_{j=1}^K \exp(s_j/\tau)}$

Fuse domain features:

$F_{\rm fused} = \sum_{k=1}^K w_k F_{\rm SDM}^k$

Frequency dynamic generation via learned low-pass filter and complement:

$F_{LP}(X) = F^L \ast X, \qquad F_{HP}(X) = X - F_{LP}(X)$

Gated fusion of spatial and frequency branches, followed by cross-attention and dynamic aggregation.

DCT-based frequency mask prediction, with Gumbel-Softmax gating.
Separate low/high-frequency streams processed by dedicated MoE blocks.
Final fusion via expert mixture adapted per input.

FFT-based decomposition:

$M_{\rm low}(u,v) = \sigma\left(\frac{T_l - r(u,v)}{\tau}\right), \quad M_{\rm high}(u,v) = \sigma\left(\frac{r(u,v) - T_h}{\tau}\right)$

Bandwise separation and Linformer-based fusion of cost volumes.

Across settings, FADF modules are implemented as functions or blocks within larger encoder–decoder or multi-branch architectures, with explicit pseudocode or algorithmic outlines provided in the literature.

A principal advantage of FADF is its capacity for adaptive cross-domain and cross-modal fusion. This manifests as:

Modality-specific wavelet or Fourier decompositions (AdaWAT, AdaIWAT, Dynamic Spectral Filtering), with fusion mappings learned to combine features in both spatial and frequency domains (Wang et al., 21 Aug 2025, Gao et al., 20 Feb 2025, Gu et al., 2023).
Adaptive domain selection via frequency similarity—for example, test-time computation of spectral similarity to choose among $K$ source domains (Wang et al., 9 Jan 2026).
Explicit handling of missing frequency components and spatial variations, such as in MRI reconstruction under arbitrary k-space undersampling (UniFS AMPL-guided fusion) (Li et al., 5 Dec 2025).
Mixture-of-experts gate networks trained to balance low-versus-high frequency content to suit task specifics or image regions (He et al., 2024).

FADF methods routinely outperform classical average-pooling, attention without frequency separation, or naive domain blending, with ablation studies demonstrating substantial boosts in quantitative metrics such as PSNR, SSIM, and MI, and qualitative improvement in edge fidelity and texture preservation.

4. Domain-Specific Applications and Empirical Results

FADF has been applied and validated across a spectrum of computer vision and signal processing tasks, including:

Domain-generalized retinal vessel segmentation: WaveRNet achieves state-of-the-art results under domain shift, outperforming SAM-based and frequency-naive methods (Wang et al., 9 Jan 2026).
Image deblurring: SFAFNet yields consistent 0.7–1.0 dB gain over spatial-only or frequency-only baselines across GoPro, HIDE, RealBlur, and DPDD (Gao et al., 20 Feb 2025).
Multimodal image fusion: AdaSFFuse surpasses fixed transform baselines on IVF, MEF, MFF, and MIF, with robust SSIM and spatial-frequency preservation (Wang et al., 21 Aug 2025).
Medical image fusion/adaptation: AdaFuse, AFDAN, and UniFS deploy FADF with frequency-guided attention and adversarial alignment, achieving top segmentation and reconstruction scores under rare annotation and domain shift (Gu et al., 2023, Li et al., 18 Dec 2025, Li et al., 5 Dec 2025).
Pan-sharpening: FAME shows strongest generalization on WorldView-II/III, GaoFen2 under QNR, PSNR, ERGAS (He et al., 2024).
Real-time stereo matching: MAFNet delivers best EPE/D1-all with efficient parameter usage (Xu et al., 4 Dec 2025).
Time series analysis: FusAD leverages dual Fourier/Wavelet fusion with adaptive denoising for multi-task learning (Zhang et al., 16 Dec 2025).
Underwater image enhancement: AQUA-Net and FUSION achieve leading PSNR/SSIM/UIQM/UCIQE with adaptive frequency modulation and attention-based fusion (Ali et al., 5 Dec 2025, Walia et al., 1 Apr 2025).

Across studies, FADF is typically a lightweight addition to existing architectures, often contributing <1% of total parameters (e.g., AQUA-Net's frequency branch at <0.6% of total), but yielding disproportionate gains.

5. Loss Functions, Training Strategies, and Ablation Studies

FADF frameworks adopt loss functions tailored to frequency preservation and adaptive fusion:

Frequency-domain loss: $L_f = \|\mathcal{F}(\hat I) - \mathcal{F}(\bar I)\|_1$ or similar, guiding frequency-feature modules (Gao et al., 20 Feb 2025).
Multi-term losses: Include SSIM, texture, intensity, and structural consistency, sometimes leveraging Fourier or wavelet differences (Gu et al., 2023, Wang et al., 21 Aug 2025).
Adversarial alignment: For domain adaptation, adversarial losses applied to amplitude spectra ensure domain-invariant frequency characteristics (Li et al., 18 Dec 2025).
Load balancing and attention regularization: Mixture-of-expert gate regularization to prevent expert collapse (He et al., 2024).

Comprehensive ablations demonstrate significant improvements when FADF modules are active; removing frequency adaptivity, replacing with fixed transforms, or using plain concatenation consistently leads to performance drops.

Paper	Task	FADF Mechanism
(Wang et al., 9 Jan 2026)	Retinal vessel segmentation	Wavelet similarity, fusion weights
(Gao et al., 20 Feb 2025)	Image deblurring	Learnable spectral split + gated fusion
(Wang et al., 21 Aug 2025)	Multimodal fusion	AdaWAT + Mamba blocks
(Li et al., 5 Dec 2025)	MRI reconstruction	Amplitude/phase fusion, mask prompts
(He et al., 2024)	Pan-sharpening	DCT mask, MoE gates, expert blending
(Xu et al., 4 Dec 2025)	Stereo matching	FFT mask, Linformer fusion
(Li et al., 18 Dec 2025)	Medical segmentation	Adversarial spectral fusion
(Ali et al., 5 Dec 2025)/(Walia et al., 1 Apr 2025)	Underwater enhancement	Frequency branch, gating, FGF
(Zhang et al., 16 Dec 2025)	Time series analysis	ASM (FFT + CWT), threshold gating

6. Design Trade-offs, Limitations, and Future Extensions

Fusion granularity: Most current FADF systems employ single-level spectral separation; deeper, multi-level wavelet or multi-scale FFT decompositions may afford finer granularity and further generalization (Wang et al., 9 Jan 2026).
Trade-off control: Temperature hyperparameters in softmax (e.g., $\tau$ ) directly regulate the "hardness" of domain selection versus blending; optimal values are task-dependent (Wang et al., 9 Jan 2026).
Frequency loss mitigation: Data-driven basis adaptation (learnable wavelet/FFT/DCT filters) consistently outperform fixed analytic transforms (Wang et al., 21 Aug 2025, Gao et al., 20 Feb 2025, He et al., 2024).
Computational efficiency: FADF modules are often computationally lightweight and parallelizable, and—in several cases—reduce or sidestep the high costs of traditional 3D convolutions or full self-attention (Xu et al., 4 Dec 2025, He et al., 23 Jun 2025).
Generalization to arbitrary domains: The capacity to interpolate among source domains/modalities or adapt to novel frequency mask patterns (e.g., arbitrary k-space undersampling) is a central strength of FADF, enabling single-model deployment over broad input variations without retraining (Li et al., 5 Dec 2025).

A plausible implication is that future research will explore multi-level spectral fusion, more sophisticated domain-attentive mechanisms, and applications beyond vision and time series (e.g., speech, biosignal processing), leveraging the robust, adaptive foundations of FADF.

7. Representative Pseudocode and Typical Workflow

The essential workflow of FADF combines domain-specific frequency decomposition, similarity or attention scoring, adaptive gating/fusion, and task-driven supervision. A canonical structure is as follows (cf. (Wang et al., 9 Jan 2026, Gao et al., 20 Feb 2025, Xu et al., 4 Dec 2025)):

Inputs: Feature map F_test, K source domain prototypes {proto_low^k}, {proto_high^k}, K domain-modulated features {F_SDM^k}, softmax temp τ
F_low_test  = W_low(F_test)
F_high_test = W_high(F_test)
v_low_test  = GAP(F_low_test)
v_high_test = GAP(F_high_test)
for k in range(K):
    sim_low_k  = cosine(v_low_test, proto_low^k)
    sim_high_k = cosine(v_high_test, proto_high^k)
    s_k        = 0.5 * (sim_low_k + sim_high_k)
w = softmax([s_1/τ, s_2/τ, ... s_K/τ])
F_fused = sum_k(w_k * F_SDM^k)
return F_fused

This exemplifies how FADF is realized in practice, with variations for wavelet, FFT, or DCT bases, different gating strategies (MoE, Linformer, cross-attention), and extensions to new application domains.