Y-Guided Local-Aware Frequency Attention Module

• The paper demonstrates that integrating global Fourier transforms with local 3x3 convolutions and Y-channel guidance significantly improves low-light image enhancement.
• It employs explicit frequency disentanglement and learnable spectral masking to separate true chrominance structures from high-frequency noise, as evidenced by improved PSNR values.
• The architecture parallels graph transformer modules by fusing structural cues with frequency filters, ensuring both global context and local detail preservation.

A Y-Guided Local-Aware Frequency Attention Module is a neural architecture component that combines global frequency-domain representations with explicit local spatial or structural filtering, modulated by guidance from a reference channel (typically luminance, denoted Y), to adaptively denoise or enhance features in low-quality input data. Introduced in the context of lightweight low-light image enhancement in YUV space, it is also theoretically analogous to structure-frequency refinement modules for graph neural networks, where "Y-guidance" metaphorically aligns with spectral or global context cues. The module targets scenarios with distinct signal and noise characteristics across channels or nodes, leveraging both global and local priors to preserve informative structural details while suppressing undesirable high-frequency noise (Yan et al., 24 Jan 2026, Zhai et al., 28 Apr 2025).

1. Theoretical Motivation

In low-light image enhancement, a key empirical insight is that the Y (luminance) channel predominantly suffers loss of low-frequency content, manifesting as structure and smoothness degradation, while the U and V (chrominance) channels retain low-frequency color information but exhibit significant high-frequency noise contamination. The design motivation for a Y-guided local-aware frequency attention module (Y-LAFA) is threefold:

• Explicit frequency disentanglement: Separate true chrominance structures from high-frequency noise by performing transformations in the Fourier domain, independently handling amplitude and phase.
• Y-guidance: Inject structural priors from the enhanced Y channel into the chrominance restoration process, leveraging its better-restored structure to stabilize color reconstruction.
• Spatial/local regularization: Enforce local spatial consistency to prevent color blotchiness and speckle artifacts, typically through a localized convolutional branch.

This approach is paralleled in graph transformers, where global (low-frequency) and local (high-frequency) behaviors are modulated via Laplacian eigendecomposition and node energy filters, and refinement masks correspond to both structural (eigenvalue) similarity and frequency content (Yan et al., 24 Jan 2026, Zhai et al., 28 Apr 2025).

2. Architectural Components

The canonical Y-LAFA architecture, as given in "Revisiting Lightweight Low-Light Image Enhancement: From a YUV Color Space Perspective" (Yan et al., 24 Jan 2026), is organized as follows:

1. Input Fusion: Fuse the raw UV feature maps ($I_{\text{low}}^{UV}$) with the enhanced Y features ($I_{\text{enh}}^{Y}$) via elementwise addition: $S = I_{\text{low}}^{UV} + I_{\text{enh}}^{Y}$.
2. Frequency Branch:
   • Apply a global 2D FFT to $S$ to decompose into amplitude ($X_A$) and phase ($X_P$).
   • Aggregate frequency features via global pooling and concatenate amplitude and phase statistics.
   • Pass through a two-layer MLP to predict channel-wise frequency attention weights $W$.
   • Learn per-channel spectral modulation mask $\hat{H}$ for both amplitude and phase.
   • Invert the FFT, reconstructing targeted spatial features $X'$.
3. Local-Aware Branch:
   • Apply a $3\times3$ depthwise convolution to $S$ to obtain $D$, enforcing local spatial coherence.
4. Feature Recombination:
   • Combine outputs via $X_{\text{LAFA}} = X' \odot W \odot D + S$, fusing frequency attention, local regularization, and input residual.

The architecture is strictly end-to-end differentiable and integrates seamlessly within multi-stage pipelines.

3. Signal Processing and Mathematical Formulation

The operation of Y-LAFA can be rigorously described by the following sequence (notation as in (Yan et al., 24 Jan 2026)):

• Fusion and Spectral Transform:

$S = I_{\text{low}}^{UV} + I_{\text{enh}}^{Y}$

$(X_A, X_P) = \operatorname{FFT}(S)$

• Spectral Feature Aggregation:

$$f = \operatorname{Concat}(\operatorname{AvgPool}(X_A), \operatorname{AvgPool}(X_P))$$

• Channel-wise Frequency Attention:

$$W = \sigma(\operatorname{FC}_2(\operatorname{ReLU}(\operatorname{FC}_1(f))))$$

• Learnable Spectral Masking:

$$X_A' = \hat{H} \odot X_A,\quad X_P' = \hat{H} \odot X_P$$

• Inverse FFT and Local Feature Extraction:

$X' = \operatorname{IFFT}(X_A', X_P')$

$D = \operatorname{DConv}_3(S)$

• Final Feature Synthesis:

$X_{\text{LAFA}} = X' \odot W \odot D + S$

This formulation ensures simultaneous frequency-based denoising, local structure preservation, and explicit Y-channel-informed guidance.

4. Principles of Local Awareness and Y-Guidance

Local awareness is implemented by depthwise convolution with a small kernel ($3\times3$), which directly enforces spatial smoothness, reduces salt-and-pepper chroma noise, and stabilizes color transitions at a fine scale. This operation does not disrupt global frequency filtering but complements it by addressing pixel and patch-level artifacts that are inadequately mitigated by spectral domain manipulations alone.

Y-guidance is realized via two mechanisms:

1. Residual Fusion: The input $S$ is a direct sum of UV and enhanced Y, immediately propagating high-frequency structural cues present in the luminance domain to the chrominance branches.
2. Attention Synthesis: Both the frequency attention $W$ and the learnable spectral mask $\hat{H}$ are functions of the fused UV+Y spectral statistics, enabling the model to learn which combined frequencies correspond to true chromatic edges versus noise, as inferred from the luminance restoration.

5. Correspondence with Graph Structure-Frequency Attention

An analogous mechanism appears in the context of graph transformers with structure-frequency refinement (Zhai et al., 28 Apr 2025). In this domain, the module consists of:

• Laplacian eigenvalue masks (local structure): encoding both global smoothness (low eigenvalues) and local irregularity (high eigenvalues).
• Per-node frequency energy calculations via graph Fourier transform, splitting signals into low/high frequency and quantifying each node's smooth versus detailed content.
• Frequency-energy filters: Cross-node matrices combining local (node) low-frequency and high-frequency strengths.
• Refinement mask $M$: Elementwise multiplication of structure and frequency-energy filters, inserted multiplicatively into Transformer attention logits.

This mechanism serves the role of "Y-guided" attention in the sense that structural and frequency cues are fused to control the allocation of attention resources, promoting both global context and local detail sensitivity.

6. Empirical Validation and Performance

Quantitative ablation in (Yan et al., 24 Jan 2026) demonstrates the contributions of each module component:

Variant	PSNR (dB)
Full LAFA	27.160
w/o LAFA	26.551
w/o Phase	26.748
w/o Amplitude	26.624
w/o Y-Guided	26.426

The performance drop associated with ablating Y-guidance, frequency or phase contributions confirms their criticality. Qualitatively, removal of amplitude or phase leads to disconnected chroma edges and increased noise; exclusion of Y-guidance results in blurred boundaries and residual artifacts.

On graph data, coupling structure and frequency in attention yields improved classification outcomes and more selective, pattern-aware attention allocation, especially under data scarcity (Zhai et al., 28 Apr 2025).

7. Integration and Applications

Y-LAFA is best employed as an intermediate component within a cascaded restoration or enhancement system:

1. Pre-enhancement: Luminance (Y) channel restoration (e.g., via Dual-Stream Global-Local Attention).
2. Chroma enhancement: Refinement of UV/chrominance channels with Y-LAFA, injecting structural priors from Y and enforcing local and global frequency regularization.
3. Fusion: Final feature merging and conversion back to RGB for output image synthesis.

A plausible implication is that the principles of frequency-structure attention, with explicit guiding signals, provide a template for designing compact, interpretable modules in tasks characterized by non-uniform channel degradations or cross-modal dependencies. This extends to both visual and non-visual modalities, including graph-structured data, where analogous signal/noise separation and structure-guided message passing are beneficial (Yan et al., 24 Jan 2026, Zhai et al., 28 Apr 2025).