Synergistic Wavelet-Attention Network (SWAN)

• SWAN is a network architecture that integrates wavelet transforms with attention mechanisms to extract both spatial and frequency information.
• It employs a staged decomposition–attention–fusion process to refine features, leveraging multi-level wavelet operators with specialized attention modules.
• Empirical evaluations across infrared detection, traffic forecasting, long-sequence learning, and image forensics validate SWAN’s capability to boost accuracy and efficiency.

Searching arXiv for SWAN and closely related wavelet-attention papers to ground the article in current literature. “Synergistic Wavelet-Attention Network” (SWAN) denotes a class of architectures that combine wavelet-domain decomposition with attention mechanisms so that feature extraction is jointly informed by spatial structure and frequency structure. In the current arXiv literature, the name is used most directly for infrared small target detection in the paper “SWAN: Synergistic Wavelet-Attention Network for Infrared Small Target Detection” (Jing et al., 2 Aug 2025), and it also appears as a descriptive label for related wavelet-attention designs in traffic forecasting, long-sequence learning, and image forensics (Fang et al., 2021, Zhuang et al., 2022, Aghdaie et al., 2021). Across these uses, the common design principle is a staged interaction between wavelet transforms, which expose multi-resolution or high-frequency content, and attention modules, which reweight spatial, temporal, or channel responses. This suggests that SWAN is best understood not as a single canonical block but as a recurrent architectural motif whose concrete instantiation is task-dependent.

1. Nomenclature and scope

The designation SWAN is used explicitly in “SWAN: Synergistic Wavelet-Attention Network for Infrared Small Target Detection” (Jing et al., 2 Aug 2025). In that work, SWAN is a U-Net-style encoder–decoder with deep supervision and three named plug-in modules: Haar Wavelet Convolution (HWConv), Shifted Spatial Attention (SSA), and Residual Dual-Channel Attention (RDCA) (Jing et al., 2 Aug 2025).

Related papers use different official names but embody the same wavelet-attention coupling. In traffic forecasting, the paper “Spatio-Temporal meets Wavelet: Disentangled Traffic Flow Forecasting via Efficient Spectral Graph Attention Network” introduces a model called STWave in the paper, but the supplied description identifies it as SWAN and centers its “key synergy” on wavelet disentangling plus graph attention (Fang et al., 2021). In long-sequence modeling, “WavSpA: Wavelet Space Attention for Boosting Transformers’ Long Sequence Learning Ability” describes “the core idea behind SWAN (Synergistic Wavelet-Attention Network), also called WavSpA in the paper” (Zhuang et al., 2022). In morph detection, “Attention Aware Wavelet-based Detection of Morphed Face Images” is likewise presented in the supplied description as a SWAN instance that combines a wavelet front-end with trainable soft attention (Aghdaie et al., 2021).

A practical consequence is that the term is polysemous. In strict bibliographic usage, SWAN most naturally refers to the IRSTD architecture of 2025 (Jing et al., 2 Aug 2025). In broader methodological usage, it refers to architectures that factor representation learning into a wavelet stage and an attention stage, often with explicit claims of synergy (Fang et al., 2021, Zhuang et al., 2022, Aghdaie et al., 2021).

2. Core architectural principle

The defining pattern is a decomposition–attention–fusion pipeline. Wavelet operators first expose structured frequency components, after which attention operates either on spatial layouts, temporal tokens, graph nodes, or channels. The resulting attended features are then fused, reconstructed, or decoded.

In the IRSTD SWAN, HWConv “jointly extract[s] spatial details and low-frequency energy patterns via a two-level nested wavelet+conv structure,” SSA “models long-range dependencies with only linear complexity,” and RDCA “adaptively calibrates channel-wise feature responses to suppress background interference while amplifying target-pertinent signals” (Jing et al., 2 Aug 2025). The top-level dataflow is given as: Input $X \to$ encoder blocks with HWConv, skip features processed by SSA, decoder blocks with RDCA, and final $1 \times 1$ convolution with deep supervision (Jing et al., 2 Aug 2025).

In STWave, the input traffic sequence $X \in \mathbb{R}^{T_1 \times N \times C}$ is first split by discrete wavelet transform into low- and high-frequency components, then processed by a dual-channel spatio-temporal encoder. The low-frequency channel uses masked temporal attention followed by efficient spectral graph attention, whereas the high-frequency channel uses one-layer dilated causal convolution followed by the same graph attention mechanism (Fang et al., 2021). The paper summary states that the model’s “key synergy comes from (a) disentangling long-/short-term patterns via wavelet DWT + dual-channel temporal modules, and (b) jointly capturing global spatial dependencies via a wavelet-positional-encoded, query-sampled spectral graph attention” (Fang et al., 2021).

In WavSpA, the attention layer itself is relocated into wavelet coefficient space. A block “replaces the standard self-attention in a Transformer layer” with three steps: forward wavelet projection, self-attention in wavelet space, and backward wavelet reconstruction (Zhuang et al., 2022). This design keeps reasoning in a multi-resolution representation while returning outputs to the original token domain.

In the morph-detection system, a three-level undecimated 2D DWT generates “48 high-frequency sub-bands,” which are stacked and passed into an Inception-ResNet-v1 backbone. Soft attention heads then compute compatibility between local feature vectors and a global feature vector, producing attentive features from multiple intermediate layers (Aghdaie et al., 2021).

A plausible implication is that the SWAN label is most coherent when reserved for architectures in which wavelet decomposition is not merely a preprocessing step but structurally coupled to attention-based reweighting.

3. Wavelet-domain components

Wavelet processing varies across SWAN variants, but each instance uses it to impose a coarse-to-fine or low-/high-frequency factorization before or during attention.

In the IRSTD model, HWConv is built from Haar-DWT filters $\phi_{LL} = \tfrac12 [1\ 1; 1\ 1]$, $\psi_{LH} = \tfrac12 [1\ -1; 1\ -1]$, $\psi_{HL} = \tfrac12 [1\ 1; -1\ -1]$, and $\psi_{HH} = \tfrac12 [1\ -1; -1\ 1]$ (Jing et al., 2 Aug 2025). For input $X \in \mathbb{R}^{C \times H \times W}$, level-1 decomposition yields $$\{F_1^{LL}, F_1^{LH}, F_1^{HL}, F_1^{HH}\} = \mathrm{DWT}(X)$$, level-2 decomposition is applied recursively to $F_1^{LL}$, each level is processed by $1 \times 1$0 convolution, and reconstruction is defined by

$1 \times 1$1

with two-level nesting in SWAN (Jing et al., 2 Aug 2025).

In STWave, the wavelet stage is a 1D discrete wavelet transform over traffic sequences. With low-pass filter $1 \times 1$2 and high-pass filter $1 \times 1$3, the model computes downsampled components $1 \times 1$4 and $1 \times 1$5, with index form $1 \times 1$6 and $1 \times 1$7 (Fang et al., 2021). To restore the original time length, inverse filters and learned lifts are applied: $1 \times 1$8 and $1 \times 1$9, with $X \in \mathbb{R}^{T_1 \times N \times C}$0 and $X \in \mathbb{R}^{T_1 \times N \times C}$1 (Fang et al., 2021).

In WavSpA, the wavelet transform is formulated through approximation and detail coefficients:

$X \in \mathbb{R}^{T_1 \times N \times C}$2

with inverse reconstruction through synthesis filters (Zhuang et al., 2022). The method explicitly allows fixed and adaptive wavelets, including AdaWavSpA, OrthoWavSpA, and LiftWavSpA (Zhuang et al., 2022).

In the morph-detection system, the front-end uses a three-level undecimated Mallat-style 2D transform and retains high-frequency sub-bands $X \in \mathbb{R}^{T_1 \times N \times C}$3, $X \in \mathbb{R}^{T_1 \times N \times C}$4, and $X \in \mathbb{R}^{T_1 \times N \times C}$5, discarding low-frequency $X \in \mathbb{R}^{T_1 \times N \times C}$6 components after level 1. The supplied description states that this yields “48 high-frequency sub-bands” of size $X \in \mathbb{R}^{T_1 \times N \times C}$7 stacked as a tensor of shape $X \in \mathbb{R}^{T_1 \times N \times C}$8 (Aghdaie et al., 2021).

These formulations differ substantially in dimensionality and intent. In IRSTD and morph detection, wavelets primarily expose image detail and target-relevant high-frequency structure (Jing et al., 2 Aug 2025, Aghdaie et al., 2021). In traffic forecasting, they separate long-term and short-term temporal schemas (Fang et al., 2021). In sequence modeling, they define the coordinate system in which attention itself is learned (Zhuang et al., 2022).

4. Attention mechanisms and their interaction with wavelets

The attention side of SWAN is equally heterogeneous. What unifies the variants is that attention receives wavelet-derived features or wavelet-space representations rather than raw signals alone.

In the IRSTD SWAN, SSA uses a windowed mechanism with cyclic shift. It comprises Window Self-Attention (WSA), in which non-overlapping $X \in \mathbb{R}^{T_1 \times N \times C}$9 windows are processed by

$\phi_{LL} = \tfrac12 [1\ 1; 1\ 1]$0

followed by Shifted-Window Attention, which applies a cyclic shift by $\phi_{LL} = \tfrac12 [1\ 1; 1\ 1]$1, WSA, and reverse shift, with a learnable relative positional bias $\phi_{LL} = \tfrac12 [1\ 1; 1\ 1]$2 (Jing et al., 2 Aug 2025). RDCA then fuses upsampled deep features and attended shallow features by channel calibration, using pooled descriptors from both streams, sigmoid scaling, and residual decoding (Jing et al., 2 Aug 2025).

In STWave, the attention mechanism is graph-based and explicitly injects wavelet positional encoding. At time step $\phi_{LL} = \tfrac12 [1\ 1; 1\ 1]$3, node features are augmented by $\phi_{LL} = \tfrac12 [1\ 1; 1\ 1]$4, where $\phi_{LL} = \tfrac12 [1\ 1; 1\ 1]$5 and $\phi_{LL} = \tfrac12 [1\ 1; 1\ 1]$6, with $\phi_{LL} = \tfrac12 [1\ 1; 1\ 1]$7 as the scale-parameterized graph wavelet filter (Fang et al., 2021). Attention is then computed with sampled queries:

$\phi_{LL} = \tfrac12 [1\ 1; 1\ 1]$8

$\phi_{LL} = \tfrac12 [1\ 1; 1\ 1]$9

The set of active query nodes has size $\psi_{LH} = \tfrac12 [1\ -1; 1\ -1]$0, reducing per-step attention from $\psi_{LH} = \tfrac12 [1\ -1; 1\ -1]$1 to $\psi_{LH} = \tfrac12 [1\ -1; 1\ -1]$2 (Fang et al., 2021).

In WavSpA, attention is standard self-attention algebraically, but it is performed on the wavelet coefficient tensor $\psi_{LH} = \tfrac12 [1\ -1; 1\ -1]$3 rather than on the original sequence:

$\psi_{LH} = \tfrac12 [1\ -1; 1\ -1]$4

$\psi_{LH} = \tfrac12 [1\ -1; 1\ -1]$5

followed by inverse wavelet reconstruction $\psi_{LH} = \tfrac12 [1\ -1; 1\ -1]$6 (Zhuang et al., 2022).

In the morph-detection network, attention is formulated as compatibility between local feature vectors $\psi_{LH} = \tfrac12 [1\ -1; 1\ -1]$7 and a global vector $\psi_{LH} = \tfrac12 [1\ -1; 1\ -1]$8:

$\psi_{LH} = \tfrac12 [1\ -1; 1\ -1]$9

and three such attentive vectors are concatenated for the final morph/bona-fide classification (Aghdaie et al., 2021).

The main misconception these papers collectively counter is that “wavelet-attention” denotes a single standard operator. The literature instead covers window self-attention, graph attention, channel calibration, and compatibility-based soft attention, all conditioned by wavelet-derived structure (Jing et al., 2 Aug 2025, Fang et al., 2021, Zhuang et al., 2022, Aghdaie et al., 2021).

5. Task-specific instantiations

The existing SWAN literature spans several domains. The table below summarizes the best-defined instances.

Variant	Domain	Defining components
SWAN (Jing et al., 2 Aug 2025)	Infrared small target detection	HWConv, SSA, RDCA, U-Net-style encoder–decoder with deep supervision
STWave / SWAN description (Fang et al., 2021)	Traffic flow forecasting	DWT disentangling, dual-channel encoder, efficient spectral graph attention
WavSpA / SWAN description (Zhuang et al., 2022)	Long-sequence learning	DWT, attention in wavelet coefficient space, IDWT reconstruction
SWAN description in morph detection (Aghdaie et al., 2021)	Morphed face image detection	Three-level undecimated 2D DWT, stacked wavelet sub-bands, multi-head soft attention

The IRSTD SWAN is designed for “complex backgrounds” in which conventional convolutions “primarily capture local spatial patterns and struggle to distinguish the unique frequency-domain characteristics of small targets from intricate background clutter” (Jing et al., 2 Aug 2025). The architecture therefore combines frequency energy extraction, long-range spatial context, and decoder-stage channel recalibration.

The STWave system addresses traffic forecasting under “different schemas,” separating low-frequency long-term trends from high-frequency short-term bursts and coupling both streams to graph attention with wavelet-based positional encoding (Fang et al., 2021).

WavSpA targets the long-sequence limitations of Transformers, arguing that wavelet transforms are preferable to Fourier transforms because they retain both position and frequency information while admitting fast wavelet transforms with linear time complexity (Zhuang et al., 2022).

The morph-detection model exploits the hypothesis that morphing artifacts are concentrated in high-frequency residuals and specific facial regions. Its attention maps are reported to concentrate on the eyes, eyebrows, and hairline, which the authors associate with landmark-based morphing perturbations (Aghdaie et al., 2021).

A plausible implication is that SWAN architectures are particularly attractive in problems where relevant signal is sparse, high-frequency, multi-scale, or spatially localized against dominant background structure.

6. Empirical results, complexity, and limitations

For IRSTD-Real, SWAN reports Param $\psi_{HL} = \tfrac12 [1\ 1; -1\ -1]$0M and Flops $\psi_{HL} = \tfrac12 [1\ 1; -1\ -1]$1G, with $\psi_{HL} = \tfrac12 [1\ 1; -1\ -1]$2, $\psi_{HL} = \tfrac12 [1\ 1; -1\ -1]$3, $\psi_{HL} = \tfrac12 [1\ 1; -1\ -1]$4, $\psi_{HL} = \tfrac12 [1\ 1; -1\ -1]$5, and $\psi_{HL} = \tfrac12 [1\ 1; -1\ -1]$6; the paper notes that this is the best result on $\psi_{HL} = \tfrac12 [1\ 1; -1\ -1]$7, $\psi_{HL} = \tfrac12 [1\ 1; -1\ -1]$8, $\psi_{HL} = \tfrac12 [1\ 1; -1\ -1]$9, and $\psi_{HH} = \tfrac12 [1\ -1; -1\ 1]$0, while $\psi_{HH} = \tfrac12 [1\ -1; -1\ 1]$1 is second best and “0.07 behind DATransNet” (Jing et al., 2 Aug 2025). On NUDT, SWAN reports $\psi_{HH} = \tfrac12 [1\ -1; -1\ 1]$2 (2nd), $\psi_{HH} = \tfrac12 [1\ -1; -1\ 1]$3 (best), $\psi_{HH} = \tfrac12 [1\ -1; -1\ 1]$4 (best), $\psi_{HH} = \tfrac12 [1\ -1; -1\ 1]$5 (best), and $\psi_{HH} = \tfrac12 [1\ -1; -1\ 1]$6 (best) (Jing et al., 2 Aug 2025). Module-wise ablation shows a progression from UNet to +HWConv to +HWConv+SSA to full SWAN, with the full model reaching $\psi_{HH} = \tfrac12 [1\ -1; -1\ 1]$7 $\psi_{HH} = \tfrac12 [1\ -1; -1\ 1]$8, $\psi_{HH} = \tfrac12 [1\ -1; -1\ 1]$9 $X \in \mathbb{R}^{C \times H \times W}$0, $X \in \mathbb{R}^{C \times H \times W}$1 $X \in \mathbb{R}^{C \times H \times W}$2, $X \in \mathbb{R}^{C \times H \times W}$3 $X \in \mathbb{R}^{C \times H \times W}$4, and $X \in \mathbb{R}^{C \times H \times W}$5 $X \in \mathbb{R}^{C \times H \times W}$6 on IRSTD-Real / NUDT (Jing et al., 2 Aug 2025). The same study states that “two-level wavelet nesting is optimal—further levels yield diminishing returns” (Jing et al., 2 Aug 2025).

For STWave, experiments on four PeMS datasets are summarized as showing “the higher traffic forecasting precision with lower computational cost” (Fang et al., 2021). The provided metrics for PeMSD4, PeMSD7, and PeMSD8 are MAE $X \in \mathbb{R}^{C \times H \times W}$7, $X \in \mathbb{R}^{C \times H \times W}$8, and $X \in \mathbb{R}^{C \times H \times W}$9; RMSE $$\{F_1^{LL}, F_1^{LH}, F_1^{HL}, F_1^{HH}\} = \mathrm{DWT}(X)$$0, $$\{F_1^{LL}, F_1^{LH}, F_1^{HL}, F_1^{HH}\} = \mathrm{DWT}(X)$$1, and $$\{F_1^{LL}, F_1^{LH}, F_1^{HL}, F_1^{HH}\} = \mathrm{DWT}(X)$$2; and MAPE $$\{F_1^{LL}, F_1^{LH}, F_1^{HL}, F_1^{HH}\} = \mathrm{DWT}(X)$$3, $$\{F_1^{LL}, F_1^{LH}, F_1^{HL}, F_1^{HH}\} = \mathrm{DWT}(X)$$4, and $$\{F_1^{LL}, F_1^{LH}, F_1^{HL}, F_1^{HH}\} = \mathrm{DWT}(X)$$5 (Fang et al., 2021). The description states that these “beat the previous best (AGCRN) by roughly 1.3 MAE / 1.8 RMSE / 0.5–1.2 % MAPE on each dataset” (Fang et al., 2021). Complexity is reduced from

$$\{F_1^{LL}, F_1^{LH}, F_1^{HL}, F_1^{HH}\} = \mathrm{DWT}(X)$$6

to $$\{F_1^{LL}, F_1^{LH}, F_1^{HL}, F_1^{HH}\} = \mathrm{DWT}(X)$$7, and the “Full” variant without query sampling is reported to run “~ 20 % faster” with “~ 30 % less GPU memory” in favor of SWAN (Fang et al., 2021).

For WavSpA on Long Range Arena, the baseline Transformer has mean test accuracy $$\{F_1^{LL}, F_1^{LH}, F_1^{HL}, F_1^{HH}\} = \mathrm{DWT}(X)$$8, fixed Daubechies-2 WavSpA boosts full-attention to $$\{F_1^{LL}, F_1^{LH}, F_1^{HL}, F_1^{HH}\} = \mathrm{DWT}(X)$$9 on Text and improves $F_1^{LL}$0 tasks, and AdaWavSpA + full attention achieves mean $F_1^{LL}$1 without Retrieval versus $F_1^{LL}$2 baseline (Zhuang et al., 2022). The same summary states that WavSpA combined with Linformer, Performer, LinearAttention, and Longformer “uniformly improves each architecture by 5–20 points” (Zhuang et al., 2022). Overhead is described as linear in sequence length, with fixed-wavelet WavSpA adding “≈3%–5% runtime,” adaptive schemes “10%–20%,” and lifting “can even speed up due to shorter subbands” (Zhuang et al., 2022).

For morph detection, SWAN reports D-EER / BPCER@APCER=5% / BPCER@APCER=10% of $F_1^{LL}$3 on VISAPP17, $F_1^{LL}$4 on LMA, and $F_1^{LL}$5 on MorGAN in single-dataset training (Aghdaie et al., 2021). Under universal training, it reports $F_1^{LL}$6 on VISAPP17, $F_1^{LL}$7 on LMA, $F_1^{LL}$8 on MorGAN, and $F_1^{LL}$9 on the universal test (Aghdaie et al., 2021). Ablation further shows that moving from one attention head to three reduces D-EER on LMA from $1 \times 1$00 to $1 \times 1$01, and on Universal$1 \times 1$02LMA from $1 \times 1$03 to $1 \times 1$04 (Aghdaie et al., 2021).

The limitations stated in the sources are likewise domain-specific. The IRSTD SWAN notes that “two-level wavelet nesting is optimal—further levels yield diminishing returns” and that “SSA adds a small overhead in highly synthetic data (slight Fa increase)” (Jing et al., 2 Aug 2025). WavSpA identifies extra parameters and modest runtime overhead for adaptive parameterizations, larger compute for OrthoWavSpA as filter length grows, an unexplored lifting-scheme design space, and an open challenge in choosing the best wavelet family by layer and modality (Zhuang et al., 2022).

7. Relation to adjacent wavelet-attention literature

SWAN belongs to a wider family of wavelet-driven attention networks in contemporary vision and sequence modeling.

For weak-boundary polyp detection, “MEGANet-W: A Wavelet-Driven Edge-Guided Attention Framework for Weak Boundary Polyp Detection” introduces a two-level Haar wavelet head and Wavelet Edge Guided Attention modules that inject “directional, parameter free Haar wavelet edge maps into each decoder stage to recalibrate semantic features” (Tan, 3 Jul 2025). On five public polyp datasets, the abstract reports that MEGANet-W improves “mIoU by up to 2.3% and mDice by 1.2%, while introducing no additional learnable parameters” (Tan, 3 Jul 2025). This is not named SWAN in the paper title, but it clearly exemplifies the same coupling of wavelet edges and attention-guided feature refinement.

For channel attention, “WaveNets: Wavelet Channel Attention Networks” replaces Global Average Pooling with wavelet transform compression and states that “global average pooling is equivalent to the recursive approximate Haar wavelet transform” (Salman et al., 2022). On ImageNet with ResNet-34, WaveNet-C-34 reports Top-1 $1 \times 1$05 and Top-5 $1 \times 1$06, compared with SENet-34 at Top-1 $1 \times 1$07 and Top-5 $1 \times 1$08, with “parameter increase = 0” and “compute overhead +0.05%” (Salman et al., 2022). This line of work indicates that wavelet-attention synergy can be implemented not only through self-attention or spatial attention but also through richer channel descriptors.

Taken together, these neighboring results support a broader interpretation of SWAN as an architectural family in which wavelet transforms serve as structured compressors, disentanglers, or edge extractors, while attention performs selective routing, contextual aggregation, or calibration (Tan, 3 Jul 2025, Salman et al., 2022). At the same time, the literature does not yet define a single standardized SWAN block across domains. The strongest precise usage remains the IRSTD model of 2025 (Jing et al., 2 Aug 2025), whereas the more general meaning is methodological rather than taxonomic.