Directional Frequency-Spatial Encoder (DFSE)

• Directional Frequency-Spatial Encoder (DFSE) is a module that jointly models orientation, frequency-domain structure, and spatial organization to capture anisotropic signal patterns.
• It integrates wavelet-based processing, directional difference convolutions, and dual-resolution context through its submodules FSE, DAC, and DRU to enhance performance efficiently.
• Empirical results in demoiréing and other domains demonstrate DFSE’s effectiveness in boosting PSNR and SSIM while reducing model parameters compared to conventional methods.

Searching arXiv for the cited DFSE-related papers and closely related work to ground the article in current literature. The Directional Frequency-Spatial Encoder (DFSE) is a feature-encoding module that jointly models directionality, frequency-domain structure, and spatial organization. In the supplied literature, the term is used most concretely in MoiréNet, where DFSE is the first main encoder block in a compact U-Net for image demoiréing and is designed to encode anisotropic, periodic moiré artifacts through wavelet-based frequency processing, directional difference convolution, and dual-resolution residual context (Guo et al., 23 Sep 2025). More broadly, the same phrase can describe a class of modules that combine directional priors with frequency-spatial fusion, as in structured medical-image decoders that integrate directional attention and Fourier/wavelet branches (Zhang et al., 5 Dec 2025), or more abstractly, encoders whose coefficients are indexed by position, scale, orientation, and frequency in multiscale geometric representation theory (Jacques et al., 2011). In this sense, DFSE denotes both a specific architectural block and a wider representational principle.

1. Conceptual basis and scope

DFSE arises from the observation that many signals exhibit structure that is simultaneously oriented, spectral, and localized. In the demoiréing setting, moiré patterns are described as originating from spectral aliasing between display pixels and sensor grids and as being anisotropic, multi-scale, periodic, and strongly coupled to genuine image textures (Guo et al., 23 Sep 2025). This motivates an encoder that does not treat all directions or all frequencies uniformly.

The specific DFSE in MoiréNet is therefore designed to do three things at once: jointly process spatial and frequency information, explicitly encode directionality through specialized difference convolutions, and model both local detail and long-range context (Guo et al., 23 Sep 2025). The same design logic appears in other domains. In medical image segmentation, a decoder framework combines learnable directional guidance with fusion across spatial, Fourier, and wavelet domains; although that paper does not name the block DFSE, it explicitly motivates a combined directional and frequency-spatial transformation (Zhang et al., 5 Dec 2025). In a different modality, a multilingual ASR model applies self-attention along the frequency direction of Mel-filterbank features, effectively implementing a directional encoder over spectral bins and suggesting a natural extension toward a fuller directional frequency-spatial encoder (Dobashi et al., 2022).

From a broader theoretical perspective, DFSE belongs to the lineage of multiscale geometric representations that “intertwine spatial, directional and frequency selectivity,” where signals are encoded by atoms localized in space, tuned to orientation, and concentrated in specific frequency regions (Jacques et al., 2011). This suggests that the modern neural DFSE can be read as a learned instantiation of an older signal-representation principle.

2. Canonical instantiation in MoiréNet

In MoiréNet, DFSE is the first main encoder block in a U-Net-style demoiréing architecture (Guo et al., 23 Sep 2025). The network takes an input image $I \in \mathbb{R}^{H \times W \times 3}$, performs shallow feature extraction, and then applies DFSE before later encoder stages, a bottleneck Frequency-Spatial Adaptive Selector (FSAS), and a decoder with pixel-shuffle and transposed convolutions. The final reconstruction follows residual learning: $I_{\text{demoiré}} = I + F(I).$

Within this system, DFSE is presented as the specialized front-end responsible for encoding moiré orientation and fine periodic structure (Guo et al., 23 Sep 2025). Its position at the entrance of the encoder path is significant: it processes shallow features before deeper multi-scale representations are formed, thereby injecting directional and frequency-aware priors early into the hierarchy.

The DFSE block is composed of three chained submodules:

Submodule	Role	Key mechanism
FSE	Dual-domain encoding	Spatial depth-wise convolution plus wavelet-domain processing
DRU	Context aggregation	Dual-resolution residual processing
DAC	Directional encoding	Multi-branch directional difference convolution

This composition distinguishes DFSE from a conventional convolutional encoder block. Rather than using only isotropic local convolutions, it combines a spatial branch with wavelet decomposition and then augments the result with direction-specific operators and cross-scale residual context (Guo et al., 23 Sep 2025).

The paper states that DFSE is one of the main reasons MoiréNet achieves strong restoration quality with only 5.513M parameters, and it reports that this parameter count represents a 48% reduction compared to ESDNet-L (Guo et al., 23 Sep 2025). The article’s technical significance therefore lies not only in the definition of DFSE, but in its role as a compact inductive bias tailored to anisotropic aliasing artifacts.

3. Internal composition and mathematical form

The first component inside DFSE is the Frequency-Spatial Encoder (FSE), a dual-branch module operating on a feature map $F \in \mathbb{R}^{H \times W \times C}$. Its formulation is given as

$$\hat{F} = R \Big( C_d(F) + \text{IWT} \big( C_g \big( \text{DWT}(F) \big) \big) \Big), \tag{1}$$

where $C_d$ is depth-wise convolution, $\text{DWT}$ is the discrete wavelet transform, $C_g$ is group convolution on wavelet subbands, $\text{IWT}$ is inverse wavelet transform, and $R$ denotes residual refinement (Guo et al., 23 Sep 2025).

This structure gives DFSE its frequency-spatial character. The DWT decomposes features into LL, LH, HL, HH subbands, with LL corresponding to low-frequency approximation and LH/HL/HH capturing oriented high-frequency content (Guo et al., 23 Sep 2025). The supplied text states that these high-frequency subbands are where moiré-related frequencies are isolated and processed before being projected back into spatial coordinates via IWT. The spatial branch and the reconstructed frequency branch are then added and refined residually.

The second major component is the Detail Augmented Convolution (DAC), which introduces explicit directionality. DAC uses five parallel branches: standard convolution, Central Difference Convolution (CDC), Angular Difference Convolution (ADC), Horizontal Moiré Difference Convolution (HMDC), and Vertical Moiré Difference Convolution (VMDC) (Guo et al., 23 Sep 2025). Their outputs are fused through learnable scalar coefficients $\alpha_k$, yielding

$I_{\text{demoiré}} = I + F(I).$0

This is equivalent to an effective kernel

$I_{\text{demoiré}} = I + F(I).$1

The directional meaning of the branches is described qualitatively. HMDC emphasizes horizontal transitions and is suited to vertical stripes; VMDC emphasizes vertical transitions and is suited to horizontal stripes; ADC targets diagonal structure; CDC captures local contrast and edge information (Guo et al., 23 Sep 2025). The text provides prototypical kernels as conceptual aids, but it explicitly frames these as qualitative descriptions rather than canonical learned coefficients.

The third component is the Dual Residual Unit (DRU), which splits the feature tensor into a high-resolution branch and a low-resolution branch, processes the latter through downsampling, residual blocks, and upsampling, and then fuses both branches (Guo et al., 23 Sep 2025). The supplied details characterize this as a “U-Net-in-a-block” design that injects long-range and multi-scale context into DFSE’s encoded features.

Taken together, these three parts establish DFSE as a hybrid encoder in which wavelet-domain decomposition, orientation-sensitive filtering, and dual-resolution context are not separate stages of an overall network but are integrated inside a single encoder block (Guo et al., 23 Sep 2025).

4. Directionality, orientation encoding, and relation to classical theory

A central property of DFSE is that it encodes orientation implicitly, rather than estimating a closed-form angle $I_{\text{demoiré}} = I + F(I).$2. In MoiréNet, the orientation signal emerges through branch responses in DAC and through the oriented high-frequency subbands produced by wavelet decomposition (Guo et al., 23 Sep 2025). Regions with strong vertical moiré produce strong responses in one directional branch, while diagonal or horizontal artifacts activate others. The learnable fusion coefficients $I_{\text{demoiré}} = I + F(I).$3 then determine which orientation-sensitive responses are emphasized.

This neural mechanism has a close conceptual relationship to classical multiscale geometric representations. The panorama of oriented transforms describes representations in which an image $I_{\text{demoiré}} = I + F(I).$4 is decomposed over atoms $I_{\text{demoiré}} = I + F(I).$5 indexed by variables such as scale, orientation, and position, via expansions of the form

$I_{\text{demoiré}} = I + F(I).$6

or, for directional wavelets,

$I_{\text{demoiré}} = I + F(I).$7

(Jacques et al., 2011).

The classical literature emphasizes three intertwined properties: spatial localization, directional selectivity, and frequency selectivity (Jacques et al., 2011). Curvelets, shearlets, contourlets, dual-tree complex wavelets, and directional wavelet frames all instantiate this principle with different compromises among redundancy, orientation resolution, and computational complexity. A plausible implication is that DFSE can be interpreted as a learned, task-specific analogue of these transforms: instead of predefining a basis or frame over $I_{\text{demoiré}} = I + F(I).$8, it uses convolutional and wavelet operators whose parameters are optimized end-to-end for a target inverse problem.

This analogy is particularly clear for moiré suppression. Moiré artifacts are anisotropic oscillatory structures, and the classical theory treats such patterns as requiring joint orientation and spectral localization. DFSE operationalizes that requirement in a neural setting by embedding orientation-sensitive and wavelet-domain operators directly into the encoder (Guo et al., 23 Sep 2025, Jacques et al., 2011).

5. Empirical behavior and measured contribution

The supplied ablation study on FHDMi quantifies the incremental effect of DFSE’s internal components (Guo et al., 23 Sep 2025). The reported values are:

Configuration	PSNR	SSIM
Baseline FSE (no DAC, no DRU)	21.41	0.7896
FSE + DAC	22.51	0.8062
FSE + DAC + DRU (full DFSE)	22.99	0.8091

These values support two concrete claims stated in the data. First, adding DAC to FSE improves PSNR by +1.10 dB and SSIM by +0.0166, indicating that explicit directional encoding materially improves suppression of moiré patterns (Guo et al., 23 Sep 2025). Second, adding DRU on top of that yields a further +0.48 dB and +0.0029 SSIM, indicating that long-range context contributes beyond local directional filtering (Guo et al., 23 Sep 2025).

At the level of the full network, MoiréNet reports PSNR 22.77, SSIM 0.8078 on UHDM and PSNR 25.49, SSIM 0.8506 on FHDMi, with 5.513M parameters (Guo et al., 23 Sep 2025). The text explicitly characterizes DFSE as a key contributor to this performance-efficiency trade-off.

A related but distinct empirical result appears in multilingual ASR. There, a Transformer encoder applied along the frequency axis improves phoneme error rates across six languages, reducing the overall PER from 26.6 to 21.3, an average improvement of 5.3 points, while using fewer parameters than the baseline (Dobashi et al., 2022). Although that paper uses the term “frequency-directional attention model” rather than DFSE, the supplied detail explicitly interprets it as a Directional Frequency Encoder and notes its relevance to a more general Directional Frequency-Spatial Encoder (Dobashi et al., 2022). This suggests that the DFSE principle is not limited to image restoration.

6. Variants, reinterpretations, and cross-domain generalizations

The term DFSE is used explicitly in MoiréNet, but the provided literature also supports broader reinterpretations.

In medical image segmentation, a decoder framework introduces Adaptive Cross-Fusion Attention (ACFA), Triple Feature Fusion Attention (TFFA), and Structural-aware Multi-scale Masking Module (SMMM) (Zhang et al., 5 Dec 2025). ACFA introduces learnable guidance in planar, horizontal, and vertical directions, while TFFA fuses Spatial, Fourier, and Wavelet branches. The supplied explanation states that the paper does not use the phrase “Directional Frequency-Spatial Encoder,” but that ACFA and TFFA together can be viewed as defining one (Zhang et al., 5 Dec 2025). Formally, the directional branch produces

$I_{\text{demoiré}} = I + F(I).$9

while TFFA fuses branch outputs through attention-weighted combination of spatial, Fourier, and wavelet representations (Zhang et al., 5 Dec 2025).

This suggests a useful typology. One can distinguish:

• Block-level DFSE: a named module, as in MoiréNet, where the encoder block itself fuses directional and frequency-spatial operators (Guo et al., 23 Sep 2025).
• Compositional DFSE: a conceptual composition of directional attention and frequency-spatial fusion, even if implemented in a decoder rather than an encoder (Zhang et al., 5 Dec 2025).
• Axis-specific DFSE: a partial instance in which one axis, such as frequency in speech, is modeled directionally, leaving time or space to other components (Dobashi et al., 2022).

The flow-matching work on DEFAR introduces a different kind of generalization. There, exposure bias is decomposed into a directional correction signal and a frequency-aware spatial weighting signal (Huang et al., 26 Jun 2026). The paper names its components Anti-Drift Rectification (ADR) and Frequency Compensation (FC), and the supplied explanation explicitly interprets DEFAR as a “Directional Frequency-Spatial Encoder for exposure bias” (Huang et al., 26 Jun 2026). This is not an encoder block in the conventional CNN sense, but it encodes drift into direction and spatial-frequency weights, showing that the DFSE concept can migrate from feature extraction to training-time rectification.

7. Interpretation, limitations, and recurring misconceptions

One common misconception is that DFSE denotes a single universally standardized architecture. The supplied evidence does not support that view. The term has a concrete architectural meaning in MoiréNet (Guo et al., 23 Sep 2025), but the broader literature supports its use as a design pattern rather than a closed specification. This suggests that “DFSE” is best treated as a family resemblance concept: modules qualify insofar as they jointly encode directional, frequency, and spatial structure.

A second misconception is that DFSE requires explicit estimation of angles or orientation labels. In the MoiréNet formulation, directionality is encoded implicitly through learned responses of CDC, ADC, HMDC, and VMDC branches and through wavelet subband structure; no closed-form angle estimator is introduced (Guo et al., 23 Sep 2025). Likewise, in the medical segmentation reinterpretation, direction is represented by learnable tensors over planar, horizontal, and vertical axes rather than by analytic orientation estimation (Zhang et al., 5 Dec 2025).

A third misconception is that DFSE is inherently tied to images. The multilingual ASR paper shows a frequency-directional encoder acting on Mel bins at each time frame, and the supplied detail identifies it as a natural building block for a more general DFSE (Dobashi et al., 2022). The multiscale-geometry survey further broadens the scope to spheres, manifolds, and graphs through directional and frequency-selective frames defined on non-Euclidean domains (Jacques et al., 2011). A plausible implication is that DFSE should be understood modality-agnostically: it is a representational strategy applicable wherever signals have localized, oriented, and spectral structure.

Finally, computational overhead remains an objective concern. The medical segmentation framework reports a rise from 25.07M params, 11.85 GMac for the baseline to 42.52M params, 18.29 GMac for the full decoder with all modules (Zhang et al., 5 Dec 2025). Although MoiréNet presents DFSE as compact (Guo et al., 23 Sep 2025), the broader lesson is that directional and multi-domain fusion can be efficient or expensive depending on how the block is instantiated.

In contemporary usage, the Directional Frequency-Spatial Encoder therefore names an architectural principle with one prominent explicit realization in image demoiréing: encode features so that orientation, frequency composition, and spatial layout are modeled jointly rather than separately. Across demoiréing, medical segmentation, speech recognition, and even generative-model rectification, the recurring idea is the same: signals often reveal their most discriminative structure only when direction, spectrum, and locality are treated as coupled variables rather than independent ones (Guo et al., 23 Sep 2025, Zhang et al., 5 Dec 2025, Dobashi et al., 2022, Huang et al., 26 Jun 2026, Jacques et al., 2011).