Half Wavelet Attention Block (HWAB)

• HWAB is an architectural component that combines wavelet-domain analysis with attention mechanisms to enhance feature representation in CNNs.
• It employs dual-attention strategies—channel and spatial—to selectively modulate frequency components from discrete wavelet transforms.
• HWAB demonstrates improved image classification and low-light enhancement performance with low computational cost compared to traditional modules.

A Half Wavelet Attention Block (HWAB) is an architectural component for deep neural networks, designed to enhance feature representation by combining wavelet-domain analysis with attention mechanisms. First introduced for image classification and subsequently extended for low-light image enhancement, HWAB leverages discrete wavelet transforms (DWT) to extract and selectively emphasize frequency-based information using attention strategies, while minimizing computational overhead by processing only a portion of the input channels in the wavelet domain (Fan et al., 2022, Xiangyu, 2022).

1. Architectural Structure and Variants

HWAB appears in multiple instantiations, notably as a replacement for stride-2 convolutional blocks in CNNs and as the central residual unit in hierarchical U-Net-style models. In M-Net+ for low-light image enhancement, each multi-resolution stage replaces conventional residual blocks with HWABs, integrating them as follows:

• The input feature map $f_{\rm in} \in \mathbb{R}^{C\times H\times W}$ is split equally along the channel dimension into the identity path $f_{\rm identity}$ and the wavelet path $f_t$ (both in $\mathbb{R}^{\frac{C}{2}\times H\times W}$).
• The wavelet path processes $f_t$ through DWT, dual attention, and inverse wavelet transform (IWT).
• The identity and wavelet paths are concatenated and passed through a $3\times 3$ convolution followed by PReLU activation, after which a $1\times 1$ shortcut connection is added to form the output $f_{\rm out} \in \mathbb{R}^{C\times H\times W}$ (Fan et al., 2022).
• Alternative variants in classification models replace stride-2 convolutions, operating on the full channel set and downsampling the spatial dimensions by a factor of two (Xiangyu, 2022).

2. Wavelet Transform and Feature Decomposition

HWAB harnesses DWT for spatial-frequency decomposition. Given filters $g$ (low-pass) and $h$ (high-pass)—often Haar or orthogonal/biorthogonal wavelets—the DWT is performed channel-wise:

$f_{\rm identity}$0

$f_{\rm identity}$1 captures the coarse approximation (low-frequency), while $f_{\rm identity}$2, $f_{\rm identity}$3, and $f_{\rm identity}$4 encode horizontal, vertical, and diagonal details (high-frequency components). For the classification HWAB, only $f_{\rm identity}$5, $f_{\rm identity}$6, and $f_{\rm identity}$7 are retained; $f_{\rm identity}$8 is discarded due to noise sensitivity (Xiangyu, 2022). The wavelet path thus emphasizes selective frequency responses, enabling better preservation of structural and textural information.

3. Attention Mechanisms in the Wavelet Domain

Two principal attention strategies are used within HWAB:

• Dual-Attention Unit (DAU) for Enhancement (Fan et al., 2022):
  • Channel Attention: Applies global average pooling, two linear transformations with a reduction ratio $f_t$0, ReLU, and sigmoid to yield $f_t$1, broadcast spatially.
  • Spatial Attention: Aggregates channels via convolution, uses ReLU, applies a second convolution, and passes through a sigmoid to produce $f_t$2.
  • The two attentions are fused multiplicatively and used to modulate $f_t$3, ultimately producing $f_t$4.
• Spatial Softmax Attention for Classification (Xiangyu, 2022): High-frequency detail maps are fused $f_t$5 and used to generate per-channel spatial softmax attention $f_t$6. The low-frequency map $f_t$7 is modulated as $f_t$8, and the result is combined residually: $f_t$9.

In both cases, attention enables selective emphasis of salient features, with enhancement models employing more expressive channel-spatial mechanisms.

4. Integration, Nonlinearities, and Data Flow

The complete HWAB workflow in M-Net+ proceeds as follows (Fan et al., 2022):

1. Split $\mathbb{R}^{\frac{C}{2}\times H\times W}$0 into $\mathbb{R}^{\frac{C}{2}\times H\times W}$1 (identity path) and $\mathbb{R}^{\frac{C}{2}\times H\times W}$2 (wavelet path).
2. Compute DWT on $\mathbb{R}^{\frac{C}{2}\times H\times W}$3 to obtain $\mathbb{R}^{\frac{C}{2}\times H\times W}$4.
3. Apply DAU attention in the wavelet domain to produce $\mathbb{R}^{\frac{C}{2}\times H\times W}$5.
4. Reconstruct $\mathbb{R}^{\frac{C}{2}\times H\times W}$6.
5. Concatenate $\mathbb{R}^{\frac{C}{2}\times H\times W}$7 with $\mathbb{R}^{\frac{C}{2}\times H\times W}$8 to form $\mathbb{R}^{\frac{C}{2}\times H\times W}$9.
6. Apply $f_t$0 convolution and PReLU activation to $f_t$1, yielding $f_t$2.
7. Compute $f_t$3 convolution on $f_t$4 to produce a shortcut $f_t$5.
8. Output $f_t$6.

Nonlinearities include ReLU in the attention unit, PReLU after the main $f_t$7 conv, and sigmoid activations in attention. Batch normalization is not used within the DAU (Fan et al., 2022).

In classification settings, the block is lighter: DWT, detail-based spatial attention, modulation of low-frequency, and residual addition, with optional $f_t$8 conv for channel adjustment (Xiangyu, 2022).

5. Applications and Empirical Performance

HWAB has been demonstrated as effective in two principal domains:

• Low-Light Image Enhancement: In the HWMNet architecture, deployment of HWAB within M-Net+ yields competitive results on LOL (PSNR/SSIM/LPIPS = 24.24/0.85/0.12) and MIT-Adobe FiveK datasets (PSNR 24.44, SSIM 0.914). Restored images exhibit enhanced structural details, less over-enhancement, and improved rendition of dark regions, outperforming baseline spatial and classical methods in both quantitative and visual metrics (Fan et al., 2022).
• Image Classification: HWAB improves accuracy as a drop-in module in major CNN backbones (e.g., MobileNetV2, VGG16bn, ResNet18/34). On CIFAR-10/100, gains of up to +1.5% Top-1 accuracy over standard networks and other attention modules are observed. Biorthogonal wavelets (bior2.2, db3) yield optimal performance (Xiangyu, 2022).

Specific experiments reveal that inserting HWAB in the deeper downsampling layers maximizes benefit. HWAB achieves its improvements with zero extra weights (classification variant), and the enhancement variant maintains state-of-the-art computational efficiency (0.92T FLOPs on $f_t$9 images).

6. Implementation Considerations and Hyperparameters

Key implementation notes include:

• Wavelet Filters: Haar, biorthogonal (e.g., bior2.2, bior3.3), and Daubechies filters are supported. Short support and symmetry are found to provide the best trade-off; bior2.2 excels on CIFAR-10 and db3 on CIFAR-100 (Xiangyu, 2022).
• Attention Reduction Ratio: The channel attention branch in enhancement models uses a reduction ratio $3\times 3$0 for dimensionality reduction.
• Downsampling: The block reduces spatial dimensions by $3\times 3$1; in classification, HWAB replaces stride-2 convolutions, keeping subsequent strides at 1.
• Parameter Count: The HWAB in classification introduces no new learnable parameters except for an optional $3\times 3$2 conv; enhancement variant introduces parameters via the dual attention unit.
• Computational Cost: Only half the channels undergo DWT/IWT in enhancement, preserving efficiency.

A summary of block configuration for the main HWAB branches follows:

Variant	Channel Split	Attention Type	Learnable Weights
Enhancement (HWMNet)	Yes	Channel+Spatial	DAU weights
Classification	No	Spatial only	None (WA-1: $3\times 3$3 conv)

7. Significance, Limitations, and Comparative Analysis

The HWAB design introduces two principal benefits:

• Frequency-Aware Attention: Exploits frequency domain phase and magnitude differences to preserve edges, textures, and illumination cues, avoiding the spatial smearing common with standard convolutional blocks. Dual attention permits selective amplification of meaningful subbands and suppression of noise/artifacts (Fan et al., 2022).
• Lightweight, Modular Integration: HWAB can serve as a direct replacement for stride-2 convolutions or residual blocks, requires minimal parameter adjustment, and can be implemented efficiently via depthwise convolutions for DWT.

HWAB consistently matches or outperforms Squeeze-and-Excitation, Efficient Channel Attention, CBAM, and global-context modules in both qualitative and quantitative evaluations (Xiangyu, 2022). A plausible implication is that frequency-domain–driven attention generalizes well across restoration and recognition settings, supplanting spatial-only paradigms for content structure preservation.

Limitations highlighted by ablation studies include reduced returns when HWAB is inserted in every downsampling stage or when less-symmetric wavelets are applied. Additionally, in models without careful tuning of filter types and placement, performance gains may saturate or even degrade.

• "Half Wavelet Attention on M-Net+ for Low-Light Image Enhancement" (Fan et al., 2022)
• "Wavelet-Attention CNN for Image Classification" (Xiangyu, 2022)