Dual-Path Edge Network

• Dual-Path Edge Network is an architectural pattern that uses two specialized processing paths to preserve and align complementary information.
• It integrates explicit fusion modules to combine features from different modalities, enhancing tasks like segmentation, deraining, and real-time inference.
• The approach balances efficiency and precision by managing trade-offs between local detail and global context across vision, communications, and optimization domains.

Dual-Path Edge Network denotes a broad architectural pattern in which two coordinated processing paths are maintained so that complementary information can be preserved, aligned, and fused before a final prediction, control decision, or systems action is made. In the cited literature, the two paths may correspond to original and augmented images, semantic and boundary cues, spatial and frequency domains, high- and low-resolution streams, time and frequency attention, page-cache and NVMe-direct storage residency, or edge-level dual variables and path-level flows; accordingly, the word edge refers variously to boundary information, edge-device deployment, and edge-centric optimization rather than to a single canonical model family (Kazerouni et al., 2023, Zuo et al., 25 Sep 2025, Shi et al., 2021, Jeong et al., 29 Apr 2026, Liu et al., 30 Jun 2025).

1. Conceptual scope

The expression is not standardized across arXiv literature, but the recurring idea is stable: a model separates two information channels that would otherwise interfere if fused too early, then reconnects them through a constrained fusion mechanism. In vision, one path often preserves fine boundary structure while the other carries semantic context; in efficient inference and sensing, one path captures local or high-frequency detail while the other models global structure or low-cost execution; in networking and optimization, one path can represent edge-level state while another represents path-level or task-level outcomes (Kazerouni et al., 2023, Zuo et al., 25 Sep 2025, Zhang, 20 Jan 2026, Liu et al., 30 Jun 2025).

This broad usage produces three major interpretations. First, edge-aware dual-path models explicitly sharpen boundaries, as in FuseNet’s original/augmented self-supervised segmentation streams, DENet’s semantic BIM path and Multi-ER edge path, and CDSE-UNet’s CT-image path plus Canny-edge path (Kazerouni et al., 2023, Zuo et al., 25 Sep 2025, Ding et al., 2024). Second, edge-efficient dual-path models are designed for resource-constrained deployment, as in VDAN for CSI gesture recognition, DPNet and DDPNet for lightweight dense prediction, DP-DRSN for automatic modulation classification, D2Former for compact speech enhancement, DNA-HHE for near-network cryptography, and DUAL-BLADE for edge LLM KV-cache offloading (Zhang, 20 Jan 2026, Shi et al., 2021, Yang et al., 2020, Suman et al., 7 Jul 2025, Zhao et al., 2023, Zhao et al., 21 Dec 2025, Jeong et al., 29 Apr 2026). Third, edge-centric optimization models shift the optimization variable itself to edges or paths, as in Geminet’s edge-level dual variables and the collaborative edge inference formulation that jointly learns path selection and DNN partition (Liu et al., 30 Jun 2025, Huang et al., 2024).

This suggests that Dual-Path Edge Network is best understood as a design pattern rather than a single architecture. The common denominator is not a fixed block diagram but a decomposition principle: two paths are specialized, their interaction is explicit, and the interaction is motivated either by boundary fidelity, deployment efficiency, or structured control.

2. Recurrent architectural patterns

A first recurrent motif is the use of paired views or paired modalities. FuseNet processes an original image $X$ and an augmented image $\tilde{X}$ with a shared-weight encoder, projection block, patch embedding, and shared cross-attention, then fuses the two streams into a pixel-level clustering map $P \in \mathbb{R}^{H \times W \times K}$ (Kazerouni et al., 2023). DPCNet divides deraining into a Spatial Feature Extraction Block and a Frequency Feature Extraction Block, then couples them with an Adaptive Fusion Module rather than naive concatenation (He et al., 2024). VDAN applies the same idea to CSI, with a Subcarrier Attention Path for frequency-domain filtering and a Temporal Attention Path for temporal localization (Zhang, 20 Jan 2026). D2Former organizes complex-valued speech features into a time path and a frequency path inside both the conformer core and the encoder/decoder dual-path dilated modules (Zhao et al., 2023).

A second motif is explicit fusion machinery that preserves specialization while enabling controlled exchange. DENet uses a Bidirectional Interaction Module in which Local Self-Attention and Global Self-Attention operate on semantic and edge streams, followed by bidirectional cross-attention between them (Zuo et al., 25 Sep 2025). CDSE-UNet applies a Double SENet Feature Fusion Block so that CT features and Canny-edge features receive separate channel attention before joint fusion, and the same mechanism replaces ordinary UNet skip connections (Ding et al., 2024). DPNet keeps a High-Resolution Path and a Low-Resolution Path, then inserts Bi-direction Fusion Modules and lightweight cross-attention in the FPN to transfer detail and semantics across resolutions (Shi et al., 2021). DDPNet internalizes duality at the module level by replacing late dense layers with a Dual-Path Module composed of a standard convolution path and a dilated convolution path (Yang et al., 2020).

A third motif is asymmetry. The two paths are often deliberately unequal. In MDFA-Net, MF-Net preserves multiscale shallow information while EC-Net models continuous global trends with attention and depthwise-separable convolution (Lv et al., 16 Dec 2025). In the AMC model, the global architecture is dual-input at the signal level, with separate I/Q and A/P pipelines, while the shrinkage submodule is itself dual-path through GAP and GMP threshold estimation (Suman et al., 7 Jul 2025). In ADP for occluded person re-identification, the holistic and occluded branches share a ViT backbone but use asymmetric supervision and explicit cross-path constraints rather than identical decoding heads (Xia et al., 2023). Dual-path design therefore does not imply symmetry; it more often implies controlled imbalance.

3. Optimization and learning

Dual-path architectures frequently employ a joint objective in which a task term is supplemented by an interaction term that forces the two paths to agree where they should and to remain complementary where they should not. FuseNet makes this explicit through

$$\mathcal{L}_{\text{joint}}=\lambda_1\mathcal{L}_{ce}+\lambda_2\mathcal{L}_{CLIP}+\lambda_3\mathcal{L}_{\text{Boundary}},$$

with $\lambda_1=2.5$, $\lambda_2=0.5$, and $\lambda_3=0.5$; the three terms respectively drive clustering-based segmentation, CLIP-style alignment between original and augmented modalities, and edge-aware spatial consistency (Kazerouni et al., 2023). DPCNet combines pixel-wise $\mathcal{L}_1$, VGG-based perceptual loss, and an FFT-domain loss with $\lambda_1=1$, $\lambda_2=0.2$, and $\tilde{X}$0, thereby coupling spatial restoration with high-frequency fidelity (He et al., 2024). D2Former similarly merges TF-domain magnitude and complex losses, time-domain waveform loss, and a QNet perceptual term while jointly training a masking decoder and a spectral-mapping decoder (Zhao et al., 2023).

In boundary-aware settings, the second path is often directly supervised or regularized as an edge carrier. DENet uses BCE for edge prediction and SoftIoU for segmentation, with the total loss defined as $\tilde{X}$1 (Zuo et al., 25 Sep 2025). VDAN introduces variational posteriors $\tilde{X}$2 and $\tilde{X}$3, adds KL terms for both paths, and interprets the resulting front-end as an information-bottleneck regularizer over subcarrier and temporal attention (Zhang, 20 Jan 2026). ADP adds path-local ID and triplet losses, a global triplet loss across holistic and occluded paths, an interaction loss using cloned holistic classifier weights, and an Attention Disturbance Mask loss that adversarially increases attention to the synthetic occluder during training (Xia et al., 2023).

In systems and optimization papers, the dual-path principle is preserved but the objective changes form. Geminet does not classify or segment; it learns an iterative update rule for edge-level dual variables $\tilde{X}$4, then derives path split ratios through a Softmin over path dual costs (Liu et al., 30 Jun 2025). The collaborative edge inference paper formulates multi-path DNN partitioning as an adversarial group linear bandit with switching costs and introduces B-EXPUCB, combining blocked EXP3 for path selection and LinUCB for per-path layer assignment (Huang et al., 2024). InversionGNN trains a direct prediction path as a multi-property surrogate oracle and then uses a gradient-based Pareto search in the inversion path to update a differentiable scaffolding tree toward Pareto-optimal molecules (Niu et al., 3 Mar 2025). Across these cases, the shared structure is a coupled optimization over two paths, but the mathematical object being optimized may be a segmentation map, an attention weight distribution, a routing policy, or a molecular graph.

4. Application domains

The most visible usage is in vision and medical imaging. FuseNet treats the original and augmented image as aligned modalities for self-supervised semantic segmentation and adds boundary refinement for skin-lesion and lung datasets (Kazerouni et al., 2023). CDSE-UNet fuses Canny edge images with CT images to improve COVID-19 lesion segmentation under blurred boundaries and strong shape variability (Ding et al., 2024). DENet separates semantic/context modeling from mathematically grounded edge refinement for infrared small target detection (Zuo et al., 25 Sep 2025). DPCNet translates the same decomposition into image deraining through spatial and frequency interaction (He et al., 2024). DPNet and DDPNet adapt dual-path reasoning to efficient object detection and real-time semantic segmentation, while ADP uses holistic and occluded branches to improve occluded person re-identification (Shi et al., 2021, Yang et al., 2020, Xia et al., 2023).

A second cluster appears in wireless sensing, communications, speech, and temporal prognostics. VDAN is a lightweight preprocessing module that separately filters CSI over subcarriers and time frames before passing the refined representation to a downstream classifier (Zhang, 20 Jan 2026). The DP-DRSN AMC model uses dual representation paths for $\tilde{X}$5 and $\tilde{X}$6 and dual threshold-estimation paths through GAP and GMP inside each residual shrinkage block (Suman et al., 7 Jul 2025). D2Former uses a fully complex dual-path conformer for monaural speech enhancement, pairing time-path and frequency-path sequence modeling with dual decoders for masking and spectral mapping (Zhao et al., 2023). MDFA-Net uses MF-Net and EC-Net to combine multiscale shallow features with deep global sequence encoding for lithium-ion battery RUL prediction (Lv et al., 16 Dec 2025).

A third cluster is systems and networking. Geminet turns traffic engineering into a dual-space problem in which a lightweight network updates edge-level dual variables while path-level routing is derived analytically (Liu et al., 30 Jun 2025). The B5G eUPF paper frames MEC-versus-cloud path choice as a POMDP and uses a DQN agent to steer TEID traffic between two N6 paths, a literal dual-path edge network in the transport sense (Moreira et al., 2 May 2026). DNA-HHE exposes a dual-mode near-network accelerator in which edge-side RNS-CKKS and Rubato share a unified datapath and can be selected according to edge–cloud trade-offs (Zhao et al., 21 Dec 2025). DUAL-BLADE partitions KV-cache residency between a page-cache path and an NVMe-direct path according to runtime memory budgets (Jeong et al., 29 Apr 2026). The collaborative edge inference paper generalizes the same pattern to multiple network paths and unknown node/link parameters (Huang et al., 2024). InversionGNN shows that the idiom extends even to graph design, where the direct path predicts molecular properties and the inversion path edits structure via gradients (Niu et al., 3 Mar 2025).

5. Empirical evidence and trade-offs

Reported results vary by domain and metric, but they consistently show that the two-path decomposition is intended to improve either fidelity, robustness, or efficiency.

Model	Setting	Reported result
FuseNet (Kazerouni et al., 2023)	Self-supervised medical segmentation	PH$\tilde{X}$7: DSC $\tilde{X}$8, HM $\tilde{X}$9, XOR $P \in \mathbb{R}^{H \times W \times K}$0; Lung: DSC $P \in \mathbb{R}^{H \times W \times K}$1, HM $P \in \mathbb{R}^{H \times W \times K}$2, XOR $P \in \mathbb{R}^{H \times W \times K}$3
DENet (Zuo et al., 25 Sep 2025)	Infrared small target detection	IRSTD-1K: nIoU $P \in \mathbb{R}^{H \times W \times K}$4, mIoU $P \in \mathbb{R}^{H \times W \times K}$5, Pd $P \in \mathbb{R}^{H \times W \times K}$6, Fa $P \in \mathbb{R}^{H \times W \times K}$7; NUDT-SIRST: mIoU $P \in \mathbb{R}^{H \times W \times K}$8, Pd $P \in \mathbb{R}^{H \times W \times K}$9, Fa $$\mathcal{L}_{\text{joint}}=\lambda_1\mathcal{L}_{ce}+\lambda_2\mathcal{L}_{CLIP}+\lambda_3\mathcal{L}_{\text{Boundary}},$$0
DPNet (Shi et al., 2021)	COCO test-dev, 320×320	$$\mathcal{L}_{\text{joint}}=\lambda_1\mathcal{L}_{ce}+\lambda_2\mathcal{L}_{CLIP}+\lambda_3\mathcal{L}_{\text{Boundary}},$$1 AP, $$\mathcal{L}_{\text{joint}}=\lambda_1\mathcal{L}_{ce}+\lambda_2\mathcal{L}_{CLIP}+\lambda_3\mathcal{L}_{\text{Boundary}},$$2 GFLOPs, $$\mathcal{L}_{\text{joint}}=\lambda_1\mathcal{L}_{ce}+\lambda_2\mathcal{L}_{CLIP}+\lambda_3\mathcal{L}_{\text{Boundary}},$$3M model size
DDPNet (Yang et al., 2020)	Cityscapes test, 1024×2048	$$\mathcal{L}_{\text{joint}}=\lambda_1\mathcal{L}_{ce}+\lambda_2\mathcal{L}_{CLIP}+\lambda_3\mathcal{L}_{\text{Boundary}},$$4 mIoU with $$\mathcal{L}_{\text{joint}}=\lambda_1\mathcal{L}_{ce}+\lambda_2\mathcal{L}_{CLIP}+\lambda_3\mathcal{L}_{\text{Boundary}},$$5 FPS; $$\mathcal{L}_{\text{joint}}=\lambda_1\mathcal{L}_{ce}+\lambda_2\mathcal{L}_{CLIP}+\lambda_3\mathcal{L}_{\text{Boundary}},$$6M parameters and $$\mathcal{L}_{\text{joint}}=\lambda_1\mathcal{L}_{ce}+\lambda_2\mathcal{L}_{CLIP}+\lambda_3\mathcal{L}_{\text{Boundary}},$$7G FLOPs
VDAN (Zhang, 20 Jan 2026)	Widar3.0 gesture recognition	Baseline CNN-LSTM $$\mathcal{L}_{\text{joint}}=\lambda_1\mathcal{L}_{ce}+\lambda_2\mathcal{L}_{CLIP}+\lambda_3\mathcal{L}_{\text{Boundary}},$$8 accuracy; VDAN $$\mathcal{L}_{\text{joint}}=\lambda_1\mathcal{L}_{ce}+\lambda_2\mathcal{L}_{CLIP}+\lambda_3\mathcal{L}_{\text{Boundary}},$$9; overhead about $\lambda_1=2.5$0K parameters and $\lambda_1=2.5$1 GFLOPs
DP-DRSN (Suman et al., 7 Jul 2025)	AMC on RML datasets	$\lambda_1=2.5$2 training parameters; average accuracies $\lambda_1=2.5$3, $\lambda_1=2.5$4, and $\lambda_1=2.5$5 on RML2016.10a, RML2016.10b, and RML2018.01a
D2Former (Zhao et al., 2023)	VoiceBank+Demand speech enhancement	$\lambda_1=2.5$6M parameters; best configuration reports WB-PESQ $\lambda_1=2.5$7, CSIG $\lambda_1=2.5$8, CBAK $\lambda_1=2.5$9, COVL $\lambda_2=0.5$0, STOI $\lambda_2=0.5$1
DNA-HHE (Zhao et al., 21 Dec 2025)	Edge-side PPOC	Reduced overall latency of edge-side PPOC by $\lambda_2=0.5$2 to $\lambda_2=0.5$3
DUAL-BLADE (Jeong et al., 29 Apr 2026)	Edge LLM KV-cache offloading	Prefill and decode latency reduced by up to $\lambda_2=0.5$4 and $\lambda_2=0.5$5; SSD utilization improved by $\lambda_2=0.5$6

Ablation studies clarify where the gains come from. In FuseNet, using only $\lambda_2=0.5$7 yields PH$\lambda_2=0.5$8 DSC $\lambda_2=0.5$9, adding $\lambda_3=0.5$0 gives $\lambda_3=0.5$1, adding $\lambda_3=0.5$2 gives $\lambda_3=0.5$3, and the full joint objective reaches $\lambda_3=0.5$4, indicating complementary contributions from path alignment and edge regularization (Kazerouni et al., 2023). In CDSE-UNet, the Double SENet Feature Fusion Block outperforms simple concatenation and single-SENet fusion, and Canny gives the best DSC among Sobel, Roberts, Prewitt, and Canny variants (Ding et al., 2024).

The costs are equally consistent. DPCNet states that the dual-path design doubles forward passes per image, cross-attention and CLIP-style losses add computational overhead, and the improvement comes from extra computation for robustness and boundary quality (He et al., 2024). DPNet and DDPNet explicitly trade architectural simplicity for cross-resolution or multi-scale fusion, but do so under strict FLOP and parameter constraints (Shi et al., 2021, Yang et al., 2020). In systems work, NVMe-direct-only is not uniformly best when memory is abundant, and DNA-HHE’s edge-side acceleration does not remove cloud-side transciphering overhead (Jeong et al., 29 Apr 2026, Zhao et al., 21 Dec 2025).

6. Limitations, misconceptions, and open problems

A common misconception is that a Dual-Path Edge Network must mean two equal backbones operating in parallel. The surveyed work does not support that view. DENet is explicitly asymmetric, with BIM as a semantic/attention path and Multi-ER as a refinement path (Zuo et al., 25 Sep 2025). MDFA-Net reports that MF-Net is the stronger branch on NASA data even though the full model still improves with EC-Net added in parallel (Lv et al., 16 Dec 2025). DUAL-BLADE is dual-path at the residency level rather than the feature-extraction level, and the assignment to page-cache or NVMe-direct is dynamic rather than static (Jeong et al., 29 Apr 2026). This suggests that the defining property is coordinated specialization, not visual symmetry.

Another misconception is that the word edge always denotes image contours. Several papers do use it that way, but others use edge to mean resource-constrained edge devices or edge-level optimization variables. That semantic spread explains why some models emphasize boundary losses, while others emphasize NIC coupling, NVMe passthrough, block-level switching costs, or edge-level dual variables (Liu et al., 30 Jun 2025, Huang et al., 2024, Zhao et al., 21 Dec 2025). The concept therefore remains taxonomically open.

Open problems are stated explicitly in several sources. CDSE-UNet proposes exploring more edge detection operators and their improvements (Ding et al., 2024). MDFA-Net identifies transfer learning as a future direction for limited training data (Lv et al., 16 Dec 2025). VDAN points toward respiration monitoring and daily activity recognition as extensions of the same front-end principle (Zhang, 20 Jan 2026). Geminet suggests integrating traffic prediction into the dual-edge iterative loop or replacing the per-edge MLP with a lightweight GNN (Liu et al., 30 Jun 2025). DNA-HHE identifies cloud-side transciphering optimization as unfinished system-level work (Zhao et al., 21 Dec 2025). Taken together, these directions indicate that Dual-Path Edge Network is less a closed model class than an evolving architectural strategy for separating antagonistic objectives—local versus global, structure versus semantics, fast versus large capacity, or edge-side versus cloud-side cost—while preserving enough coupling for end-to-end optimization.