High-Frequency Enhancement Block

• High-Frequency Enhancement Block is a specialized unit that isolates, processes, and amplifies fine high-frequency components like edges and textures through frequency decomposition.
• It utilizes techniques such as Fourier, wavelet, and Laplacian filtering combined with attention, gating, and residual dense convolutions to enhance feature fidelity.
• Practical applications span super-resolution, image restoration, demosaicking, and PDE solving, with ablation studies confirming improved metrics like PSNR and SSIM.

A High-Frequency Enhancement Block (HFEB) is a specialized architectural unit—common across contemporary signal, image, and neural operator networks—designed to explicitly isolate, process, and amplify high-frequency components (fine textures, edges, oscillatory modes) that are typically lost, attenuated, or oversmoothed by standard convolutional or transformer-based processing. By decomposing features into frequency bands (via Fourier, wavelet, Laplacian, or learned high-pass operations) and applying targeted enhancement or correction modules specifically to the high-frequency bands, the HFEB enables greater fidelity in detail preservation and reconstruction throughout a wide array of tasks, including super-resolution, restoration, demosaicking, domain translation, and PDE solution.

1. Frequency Decomposition and Modality

A canonical HFEB begins by decomposing its input tensor into distinct frequency bands. This is achieved through several domain-specific strategies:

• Fourier/Laplacian/Wavelet Decomposition: Discrete Fourier Transform (DFT/FFT) is frequently used for global band separation in CNNs, SSMs, and transformer backbones, as in (Zhao et al., 2023, Cheng et al., 2024, Zhang et al., 6 Aug 2025, Chen et al., 11 May 2025). Laplacian pyramid (Gaussian smoothing + residual) provides a strict high/low split with compact support in the spatial domain, exploited in (Li et al., 2022).
• Learned High-Pass Filters: Up–downsample or 1x1 channel mixing and subtraction methods are used for approximate high-pass extraction with parametrized components (Weng et al., 2023).
• Wavelet Subband Partition: Orthogonal transforms (e.g. 2D Haar DWT) partition input features into low-frequency (LL) and three high-frequency subbands (LH, HL, HH), foundational in (Zhang et al., 2023, Azad et al., 2023, Zou et al., 2024, Zhang et al., 6 Aug 2025).

Pseudocode representing a common wavelet split for an input $F \in \mathbb{R}^{H \times W \times C}$:

h = [1/sqrt(2), 1/sqrt(2)] g = [1/sqrt(2), -1/sqrt(2)] LL = conv2d(conv2d(F, h, axis=0), h, axis=1)[::2, ::2, :] LH = conv2d(conv2d(F, h, axis=0), g, axis=1)[::2, ::2, :] HL = conv2d(conv2d(F, g, axis=0), h, axis=1)[::2, ::2, :] HH = conv2d(conv2d(F, g, axis=0), g, axis=1)[::2, ::2, :] high_freq = concat(LH, HL, HH, axis=-1)

2. Block Architecture and Enhancement Strategies

Once high-frequency features are isolated, the HFEB uses tailored parameter-efficient branches to boost or correct these components, with typical architectural features as follows:

• Amplitude/Phase Modulation in the Frequency Domain: Following FFT, separate amplitude and phase are processed by compact 1×1 or 3×3 convolutional subnets with non-linearities (LeakyReLU/GELU), after which the result is recombined and inverted back to the spatial domain (Cheng et al., 2024, Zhao et al., 2023).
• Attention and Gating: High-frequency activations are adaptively modulated via sigmoid or softmax gates derived from spatial priors (e.g., Sobel gradients), learnable masks, as well as cross-modal/multimodal attention when present. Such gates are often computed via lightweight depthwise separable or pointwise convolutions (Zhang et al., 6 Aug 2025, Weng et al., 2023).
• Residual Dense Convolutional Connectivity: Dense or skip-connected 3×3 convolutions, often with channel-wise or multi-scale fusion, reinforce locality and fine-grained propagation in the high-frequency branch (Chen et al., 2024, Weng et al., 2023).
• Guided/Multi-Path Correction: Enhanced low-frequency features can serve as guidance or input to attention/fusion operations within the high-frequency block itself (Zou et al., 2024). Some blocks employ dual-path processing, splitting bands by learned masks and applying generation/suppression modules (Liu et al., 20 Mar 2025).
• Fusion and Bottlenecking: Output features of the HF and LF branches are concatenated and passed through a final bottleneck convolution (often 1×1 or 3×3) to recombine information for the next network stage (Weng et al., 2023, Chen et al., 2024).

3. Mathematical Formulation and Data Flow

Mathematical formalism in HFEBs is exact and modular. A representative process in a patch-based DWT approach (Azad et al., 2023) is:

$$\text{DWT}(X) \to \{X_{LL}, X_{LH}, X_{HL}, X_{HH}\},\qquad X_{\mathrm{HF}} = \mathrm{concat}(X_{LH}, X_{HL}, X_{HH})$$

$$F_{HF}^{(0)} = \mathrm{ReLU}(\mathrm{Conv}_{3\times1\times1}(X_{HF}))$$

$$\forall\ i, F_{HF}^{(i+1)} = \downarrow_2 (G * F_{HF}^{(i)})$$

$$\text{Boundary\ Mask}\ B = \mathrm{Sigmoid}(\mathrm{Conv}_{1\times1}(\sum_i \mathrm{Upsample}(F_{HF}^{(i)})))$$

In attention-driven frequency matching (Zou et al., 2024), enhancement is performed by building a similarity-based map between high-frequency and enhanced low-frequency channels, then gating fused tensors:

$$M_{p,q} = \mathrm{Sim}(F_L^e[:,:,p], F_H^{\mathrm{in}}[:,:,q])$$

$Y_{\mathrm{sel}} = \mathrm{gather}_{D_q}\,F_L^e$

$$\mathrm{FMT}(F_H^{\mathrm{in}}, F_L^e) = W_p\bigl(\sigma(W_p([Y_{\mathrm{sel}}, F_H^{\mathrm{in}}])) \odot W_d([Y_{\mathrm{sel}}, F_H^{\mathrm{in}}])\bigr)$$

4. Design Variants and Task-Specific Adaptations

HFEBs are adapted per modality and task constraints:

• Image Super-Resolution: Two-branch blocks in HFFN assign explicit high-pass processing to high-frequency channels and pass through low-frequency channels with minimal computation (Weng et al., 2023).
• Image/Video Restoration, Demosaicking: Dual-path frequency enhancement as in DFENet, with learned spectrum masks selecting frequency bands for spatial detail generation or false frequency suppression (Liu et al., 20 Mar 2025).
• PDE Solving: FreqMoE's block in neural operator architectures employs expert sharing and sparse, positional gating in the Fourier domain for parameter-efficient extension to high frequencies (Chen et al., 11 May 2025).
• Underwater/Dark Image Enhancement: Gradient or structure priors from self-mined or endogenous modules are fused with wavelet high-frequency subbands for adaptive enhancement, often via spatial gating and residual fusion (Zhang et al., 6 Aug 2025, Cheng et al., 2024).
• Vision Transformers: Wavelet attention blocks inject Gaussian pyramid–derived masks computed on high-frequency bands into the Value stream of the transformer, promoting boundary-aware contextualization in self-attention (Azad et al., 2023, Zhang et al., 2023).

5. Quantitative Benefits and Ablation Analysis

HFEBs contribute significant improvements in objective metrics and visual quality over their absence, confirmed by statistical ablation:

Method	Domain	PSNR Gain (dB)	SSIM Gain	Other
HFFB in HFFN (Weng et al., 2023)	Super-Resolution	+0.33	—	31.09 on Manga109 vs. 30.48/30.53
DFF block in DSFFNet (Zhao et al., 2023)	Underwater Enhancement	+0.5–0.9	+0.0212	20.97 vs. 20.07 (full vs. no SFI)
High-Freq branch in HSIE (Li et al., 2022)	Hyperspectral LLIE	+0.929	+0.0288	SAM ↓ by –0.359 (Table VII)
HFEBlock in Wave-Mamba (Zou et al., 2024)	UHD LLIE	+0.94	+0.01	27.35 vs. 26.41 full/ablated
DFGF–HFB in SPJFNet (Zhang et al., 6 Aug 2025)	Dark Enh.	SOTA, not tabulated	—	Substantially reduced complexity
DFENet dual-path (Liu et al., 20 Mar 2025)	Demosaicking	+0.6–3+	—	32.52 vs. 29.4 on hard linset microtests

Ablation consistently shows that suppressing or omitting the specialized high-frequency block—while leaving all other architecture fixed—reduces both PSNR and visual crispness, especially in edge-rich or moiré-prone benchmarks.

6. Practical Implementation Considerations

Effective deployment of HFEBs requires attention to several factors:

• Spectral Mask Design: Use learnable, low-resolution frequency selectors (upsampled, binarized) to robustly partition Fourier bands (Liu et al., 20 Mar 2025).
• Efficiency: Depthwise or pointwise convolution and channel gating are preferred for parameter efficiency (Weng et al., 2023, Zhang et al., 6 Aug 2025).
• Guided Correction: For joint frequency–spatial methods, always recombine LF and HF features via cross-scale attention or learnable gating (Chen et al., 2024, Zou et al., 2024).
• Loss Functions: Employ multi-level, multi-domain (pixel/FFT) loss functions focusing on band-specific supervision (Liu et al., 20 Mar 2025).
• Integration: Place HFEBs after every major wavelet or pyramid split, and fuse results with corresponding low-frequency outputs, especially in U-Net or hierarchical backbones (Zou et al., 2024).
• Progressive Training: In operator learning, pretrain on low frequencies, then enable HFEB routing and supervision for improved convergence (Chen et al., 11 May 2025).

7. Applications and Generalization

HFEBs are now integral in state-of-the-art pipelines for:

• Single/multi-image super-resolution, demosaicking, and deblurring (Weng et al., 2023, Liu et al., 20 Mar 2025)
• Hyperspectral, underwater, and low-light image enhancement (Zhang et al., 6 Aug 2025, Li et al., 2022, Cheng et al., 2024)
• Neural operator learning and long-term PDE prediction (Chen et al., 11 May 2025)
• Medical segmentation and transformer models for fine-grained structure detection (Azad et al., 2023, Zhang et al., 2023)
• Multi-modal fusion and cross-modal retrieval (Zhang et al., 2023)

The underlying HFEB principle—dedicated, tuned processing of spectral high-frequency information—generalizes across neural, signal, and hybrid physical-digital optimization frameworks wherever texture, edge fidelity or high-wavenumber dynamics are critical to perceptual or quantitative success.