WMNet: Dual Wavelet-Guided Architectures

• WMNet is a homonymous term representing two wavelet-guided vision systems: one for misalignment-aware visible-infrared detection and another for HDR video reconstruction.
• The detection model employs a dual-stream YOLOv11 backbone enhanced by Wavelet U-Net and self-adaptive fusion modules to address spatial and modality misalignment.
• The HDR model uses a two-phase training strategy with Wavelet-domain Masked Image Modeling and temporal fusion modules, delivering high color fidelity and temporal consistency.

Searching arXiv for “WMNet” to disambiguate the term and verify papers. WMNet is not a single canonical architecture in the supplied arXiv literature. The name is used for at least two unrelated wavelet-centered vision systems: the “Wavelet-guided Misalignment-aware Network” for visible-infrared object detection (Zhang et al., 27 Jul 2025) and a 2026 HDR video reconstruction network introduced in “Wavelet-Domain Masked Image Modeling for Color-Consistent HDR Video Reconstruction” (Zhang et al., 7 Feb 2026). The term is also easily confused with W-net, WNSMN, and the networking abbreviation WMN (“wireless mesh network”), so precise identification requires the task, title, and arXiv id.

1. Terminological scope

In the supplied sources, “WMNet” is best treated as a homonymous label rather than a standardized architecture family. The two literal WMNet papers are both wavelet-guided, but they solve different problems, use different backbones, and report results on different benchmarks.

Usage of the name	Official designation	Task
WMNet (Zhang et al., 27 Jul 2025)	Wavelet-guided Misalignment-aware Network	Visible-infrared object detection
WMNet (Zhang et al., 7 Feb 2026)	HDR video reconstruction network with Wavelet-domain Masked Image Modeling	HDR video reconstruction

This ambiguity matters because neighboring papers in the same corpus use similar strings for entirely different systems. The retinal segmentation paper explicitly names its model W-net or $\mathcal{W}$-net, not WMNet (Zhao et al., 2020), and the numerical reasoning paper explicitly names its model WNSMN, not WMNet (Saha et al., 2021). In networking literature, WMN denotes wireless mesh network, again unrelated to either vision model (Sen et al., 2010).

2. WMNet in visible-infrared object detection

The 2025 WMNet addresses visible-infrared object detection under cross-modal misalignment. Its problem formulation is not generic multimodal fusion; it is specifically organized around Spatial Offset, Resolution Discrepancy, and Modality Deficiency / Modality Inconsistency, together with illumination variation and noise (Zhang et al., 27 Jul 2025). The architecture uses a dual-stream YOLOv11 backbone, takes the infrared modality as the reference modality, and inserts four Misalignment-Aware Fusion (MAF) modules into the backbone. Before the backbone, a Wavelet U-Net (WU-Net) enhances the visible image using the infrared image, yielding an enhanced visible input $\mathbf{I}'_{rgb}$.

WU-Net performs multilevel wavelet decomposition. For input $\mathbf{I}$, the encoder repeatedly applies a wavelet transform and splits the result into low- and high-frequency components,

$$\mathbf{F}_{0,LL} = \mathbf{I}, \qquad \mathbf{F}_{k,LL}, \mathbf{F}_{k,H} = \mathrm{Split}\!\left(\mathrm{WT}\!\left(\mathbf{F}_{k-1,LL}\right)\right).$$

At the bottleneck, frequency-aware attention (FAAttention) combines visible and infrared low- and high-frequency components. The decoder then reconstructs the enhanced visible representation through inverse wavelet transform, explicitly using RGB low-frequency structure and infrared high-frequency detail. This design is motivated by the claim that coarse structure and fine boundaries should not be fused identically under cross-modal mismatch.

The core fusion mechanism inside each MAF is Self-Adaptive Wavelet Fusion (SAWF). SAWF wavelet-decomposes the RGB feature map, applies a $3\times 3$ convolution in the wavelet domain, and modulates channels with a learnable $\mathbf{W}_1$ initialized to $0.1$, which the paper describes as reducing RGB dominance and making fusion infrared-guided. Low-frequency RGB features and infrared features are then processed by Cross-Modality Fusion Mamba (CFM). CFM adopts a Mamba2 state-space block with the standard form

$$h(t)=\mathbf{A}_t h(t-1)+\mathbf{B}_t x_t, \qquad y(t)=\mathbf{C}_t h(t)+\mathbf{D}_t x_t,$$

after pooling features to $16\times 16$, flattening them to $\mathbb{R}^{256\times C}$, and concatenating them along the sequence dimension. A wavelet-based correlated feature enhancer then reconstructs mutual information through inverse wavelet transform, sharpens object boundaries, suppresses unrelated background features, and adds a weighted RGB skip path controlled by $\mathbf{I}'_{rgb}$0. The resulting modality-specific outputs are refined by shared LayerNorm + $\mathbf{I}'_{rgb}$1 convolution, fused with same-modality skips by Scaled Add, and finally summed to form stage features $\mathbf{I}'_{rgb}$2; the last three, $\mathbf{I}'_{rgb}$3, feed the YOLO detection head.

Empirically, the detector is evaluated on DVTOD, DroneVehicle, and M$\mathbf{I}'_{rgb}$4FD. On DVTOD, the full WMNet reports $\mathbf{I}'_{rgb}$5 and $\mathbf{I}'_{rgb}$6; on DroneVehicle, $\mathbf{I}'_{rgb}$7 and $\mathbf{I}'_{rgb}$8; on M$\mathbf{I}'_{rgb}$9FD, $\mathbf{I}$0 and $\mathbf{I}$1 (Zhang et al., 27 Jul 2025). The ablation study attributes much of the gain to WU-Net and SAWF, while showing that CFM mainly improves parameter efficiency and needs the wavelet-guided misalignment machinery around it to realize its full benefit.

3. WMNet in HDR video reconstruction

The 2026 WMNet addresses a different problem: reconstruction of HDR video from LDR video sequences, with emphasis on color inaccuracies and temporal inconsistencies (Zhang et al., 7 Feb 2026). Its central idea is a two-phase training strategy. Phase I uses Wavelet-domain Masked Image Modeling (W-MIM) for self-reconstruction pre-training; Phase II fine-tunes the network for HDR video reconstruction using the pretrained encoder, a Temporal Mixture of Experts (T-MoE) module, and a Dynamic Memory Module (DMM).

The wavelet machinery is explicit. The model uses a 2D discrete wavelet transform with Haar filters

$\mathbf{I}$2

and three decomposition levels by default. W-MIM masks all high-frequency coefficients with a full-zero mask and randomly masks the deepest low-frequency component. The low-frequency masking ratio is gradually increased from 0 to 0.5 during pretraining, with mask size $\mathbf{I}$3. The masked sequence is reconstructed by an asymmetric encoder-decoder, but after Phase I only the encoder weights are retained.

Phase II adds temporal modules. T-MoE takes outputs from $\mathbf{I}$4 encoder residual groups, assigns them adaptive weights through Conv3D + Softmax, and fuses adjacent-frame information to reduce flicker. DMM provides longer-range temporal context through scene-specific memory. Each scene has its own memory space; the current feature is projected to queries, the stored scene memory to keys, and the current feature again to values, after which cross-attention is computed as

$\mathbf{I}$5

If a scene is new and has no stored memory, the cross-attention path is bypassed. The memory queue length is fixed at $\mathbf{I}$6.

The loss design is comparatively simple. Phase I uses only an $\mathbf{I}$7 reconstruction loss. Phase II uses

$\mathbf{I}$8

where $\mathbf{I}$9 is an SSIM loss. The paper does not define an adversarial loss, a perceptual VGG-style loss, or an explicit temporal-consistency loss; temporal consistency is handled architecturally by T-MoE and DMM rather than by an additional objective term (Zhang et al., 7 Feb 2026).

The main evaluation is on HDRTV4K-Scene, HDRTV4K-LongScene, and Real-HDRV. On HDRTV4K-Scene, the full model reports PSNR 36.23, SSIM 0.9630, SR-SIM 0.9987, $$\mathbf{F}_{0,LL} = \mathbf{I}, \qquad \mathbf{F}_{k,LL}, \mathbf{F}_{k,H} = \mathrm{Split}\!\left(\mathrm{WT}\!\left(\mathbf{F}_{k-1,LL}\right)\right).$$0, HDR-VDP3 8.173, LPIPS 0.0641, and $$\mathbf{F}_{0,LL} = \mathbf{I}, \qquad \mathbf{F}_{k,LL}, \mathbf{F}_{k,H} = \mathrm{Split}\!\left(\mathrm{WT}\!\left(\mathbf{F}_{k-1,LL}\right)\right).$$1. On HDRTV4K-LongScene it reports PSNR 37.19, SSIM 0.9803, and $$\mathbf{F}_{0,LL} = \mathbf{I}, \qquad \mathbf{F}_{k,LL}, \mathbf{F}_{k,H} = \mathrm{Split}\!\left(\mathrm{WT}\!\left(\mathbf{F}_{k-1,LL}\right)\right).$$2; on Real-HDRV, PSNR 27.39, SSIM 0.9507, LPIPS 0.0800, and 0.9355 ms/frame inference time (Zhang et al., 7 Feb 2026). The ablations identify W-MIM as the largest contributor, with T-MoE and DMM adding further gains.

4. Shared design patterns and sharp divergences

The two WMNet papers share a notable design preference for wavelet-domain factorization, but they operationalize it in different ways. In the detection model, wavelets are used to separate coarse structure from fine boundary detail so that cross-modal fusion can be made misalignment-aware. In the HDR model, wavelets are used to create a self-reconstruction task in which masking high-frequency coefficients disturbs color and detail information more aggressively than ordinary pixel-space masking (Zhang et al., 27 Jul 2025, Zhang et al., 7 Feb 2026).

They also share an emphasis on adaptive rather than fixed fusion. The detector uses FAAttention, Mamba-based CFM, channel-wise modulation $$\mathbf{F}_{0,LL} = \mathbf{I}, \qquad \mathbf{F}_{k,LL}, \mathbf{F}_{k,H} = \mathrm{Split}\!\left(\mathrm{WT}\!\left(\mathbf{F}_{k-1,LL}\right)\right).$$3, and Scaled Add. The HDR model uses task-weighted multi-depth temporal fusion through T-MoE and scene-specific retrieval/update through DMM. In both cases, the paper-level claim is that naive feature aggregation is insufficient: misleading information must be down-weighted, while salient structure, detail, or temporally stable context must be emphasized.

The divergences are more consequential than the commonalities. The detection WMNet is a dual-stream, multimodal detector built on YOLOv11 and targeted at RGB-infrared object detection. The HDR WMNet is an encoder-decoder video reconstruction network trained in two phases and targeted at LDR-to-HDR sequence conversion. One treats infrared as the reference modality and addresses spatial and modality misalignment; the other treats the main difficulty as color restoration plus temporal coherence. The shared name therefore does not denote a shared codebase, benchmark lineage, or standardized architecture.

5. Empirical standing

The reported results position both WMNet papers as state-of-the-art within their respective evaluations. In the detector paper, the full model outperforms strong baselines such as CMA-Det, ICAFusion, and CFT on DVTOD, and remains best on both partially misaligned DroneVehicle and aligned M$$\mathbf{F}_{0,LL} = \mathbf{I}, \qquad \mathbf{F}_{k,LL}, \mathbf{F}_{k,H} = \mathrm{Split}\!\left(\mathrm{WT}\!\left(\mathbf{F}_{k-1,LL}\right)\right).$$4FD, which the authors use as evidence that the method is not restricted to a single misalignment regime (Zhang et al., 27 Jul 2025). The same paper also reports a smaller WMNet-Lite (YOLOv11s) with 17.58M parameters and a larger WMNet (YOLOv11l) with 87.89M parameters.

In the HDR paper, the headline claim is not merely better distortion metrics but improved color fidelity, temporal coherence, and perceptual quality. The model is evaluated with seven metrics—PSNR, SSIM, SR-SIM, HDR-VDP3, $$\mathbf{F}_{0,LL} = \mathbf{I}, \qquad \mathbf{F}_{k,LL}, \mathbf{F}_{k,H} = \mathrm{Split}\!\left(\mathrm{WT}\!\left(\mathbf{F}_{k-1,LL}\right)\right).$$5, LPIPS, and $$\mathbf{F}_{0,LL} = \mathbf{I}, \qquad \mathbf{F}_{k,LL}, \mathbf{F}_{k,H} = \mathrm{Split}\!\left(\mathrm{WT}\!\left(\mathbf{F}_{k-1,LL}\right)\right).$$6—and the full WMNet is reported as best overall on HDRTV4K-Scene and highly competitive on the cross-domain Real-HDRV evaluation (Zhang et al., 7 Feb 2026). The same study reports approximately 1.36M parameters, about 48.51G FLOPs at the chosen DMM setting, and a user study favoring the method over prior baselines.

A plausible implication is that “WMNet” is currently most strongly associated with wavelet-guided selective information routing rather than with any single application domain. That implication should nevertheless be treated cautiously, because the two published systems remain task-specific and structurally incompatible.

6. Common confusions and related names

A recurring source of confusion is the proximity of “WMNet” to other names that are visually similar but technically distinct. In retinal image analysis, the official model name is W-net or $$\mathbf{F}_{0,LL} = \mathbf{I}, \qquad \mathbf{F}_{k,LL}, \mathbf{F}_{k,H} = \mathrm{Split}\!\left(\mathrm{WT}\!\left(\mathbf{F}_{k-1,LL}\right)\right).$$7-net, a shared-encoder, two-decoder segmentation network for optic disc and exudate segmentation; the supplied paper explicitly states that “WMNet does not appear” as a model name, abbreviation, alias, or variant (Zhao et al., 2020). In numerical reasoning for machine reading comprehension, the official model is WNSMN, short for Weakly Supervised Neuro-Symbolic Module Network; the supplied paper again explicitly states that the model is not literally named WMNet (Saha et al., 2021).

The term is also easy to confuse with WMN, the standard abbreviation for wireless mesh network. The supplied WMN literature covers selfish-node detection, fairness enforcement, multicast, MAC design, trust, and security, but these papers discuss networking systems rather than a model literally named WMNet (Sen et al., 2010). For bibliographic clarity, the safest practice is therefore to cite the full title and arXiv id whenever “WMNet” is mentioned.

In current arXiv usage represented here, WMNet has two precise meanings and several frequent false friends. It denotes either a wavelet-guided misalignment-aware visible-infrared detector (Zhang et al., 27 Jul 2025) or a wavelet-domain masked-image-modeling HDR video reconstruction network (Zhang et al., 7 Feb 2026); it does not, in the supplied sources, denote W-net, WNSMN, or the general WMN literature.