Dual Local-Global Encoder (DLGE)

• DLGE is an architectural paradigm that separates fine-grained local feature extraction from broad global context encoding for diverse applications.
• It employs adaptive fusion mechanisms like gated, cross-attention, or hierarchical strategies to combine complementary information effectively.
• Empirical results in domains such as histopathology, aerial imagery, and EEG decoding consistently show gains in robustness, discrimination, and generalization.

Dual Local-Global Encoder (DLGE) denotes an architectural pattern in which separate processing pathways encode fine-grained local structure and broader global context, followed by an explicit fusion stage that combines the two representations for a downstream task. In the supplied literature, the term appears directly in histopathology cell detection and cross-BCI-paradigm EEG decoding, while closely related formulations recur under names such as dual-stream, dual-encoder, dual-scale, and local-global encoder in aerial imagery, image forensics, medical segmentation, sign language retrieval, knowledge-graph reasoning, sensorless 3D ultrasound, and point-cloud analysis (Tabaghsar et al., 5 Jun 2026, Wang et al., 25 Aug 2025, Marcu et al., 2016). This suggests that DLGE is best understood as an architectural paradigm rather than a single fixed network.

1. Historical emergence and conceptual scope

The earliest work in the supplied set that clearly instantiates the local-global principle is the 2016 dual-stream model for aerial imagery, which processes information along two independent pathways, one for local and another for global visual reasoning, and combines them in the final layers for segmentation of buildings and roads (Marcu et al., 2016). Subsequent works generalized the same structural idea to image splicing forgery detection through an unfixed encoder and a fixed encoder with a spatial pyramid global-feature extraction module (Liu et al., 2020), to airborne laser scanning point clouds through local geometric extraction plus global contextual encoders (Lin et al., 2020), and to sign language retrieval through Pose and RGB streams encoding local and global information respectively (Jiang et al., 2024).

The term DLGE becomes explicit in later papers. In histopathology, DualGate-Net is described as being built on a Dual Local-Global Encoder framework in which both local cell morphology and global tissue context are critical (Tabaghsar et al., 5 Jun 2026). In EEG, DLGE is proposed as a generalizable cross-paradigm model whose local encoder learns shared within-region features and whose global encoder aggregates them into paradigm-specific representations (Wang et al., 25 Aug 2025).

Across these works, the local branch is consistently associated with high-frequency detail, short-range structure, or region-specific evidence, whereas the global branch captures long-range dependencies, tissue or scene organization, graph-scale structure, or task-level context. The literature therefore treats DLGE not as a domain-specific trick, but as a recurring response to tasks in which either local evidence alone or global context alone is insufficient.

2. Canonical architectural pattern

A DLGE system typically comprises two specialized encoders, differentiated either by input, receptive field, modality, or inductive bias. The local pathway is usually optimized for precision, boundary sensitivity, or short-range dynamics. The global pathway is usually optimized for context, long-range dependence, or coarse semantic organization.

System	Local pathway	Global pathway / task
DualGate-Net	ConvNeXtV2 local encoder for nuclear morphology and cellular boundaries	SegFormer global encoder for tissue organization; histopathology cell detection
DE-KAN	Customized CNN on original inputs	ResNet-18 on strongly augmented inputs; 2D teeth segmentation
SPG-CDENet	Local encoder on prior-masked image	Global encoder on full image; multi-organ segmentation
DualTrack	Modified 3D ResNet18 for dense spatiotemporal local cues	Image backbone plus temporal attention; sensorless 3D ultrasound

In DualGate-Net, the local encoder extracts fine-grained, high-frequency features and the global encoder captures long-range dependencies and contextual tissue organization; the two operate in parallel at matched spatial resolutions for multi-scale integration (Tabaghsar et al., 5 Jun 2026). In DE-KAN, the dual-encoder design is split between a ResNet-18 encoder for strongly augmented dental images and a customized CNN encoder for original images, explicitly targeting complementary global and local spatial features (Mustakim et al., 23 Nov 2025). In SPG-CDENet, the global encoder processes the original full image while the local encoder processes a prior-masked image derived from a spatial prior network (Tian et al., 30 Oct 2025). In DualTrack, the local encoder uses dense spatiotemporal convolutions to capture speckle patterns and frame-to-frame motion, while the global encoder uses an image backbone and temporal attention to embed anatomical features and long-range dependencies (Wilson et al., 11 Sep 2025).

The same division of labor also appears outside medical imaging. In SEDS, Pose is the local stream and RGB is the global stream (Jiang et al., 2024). In the EEG DLGE model, local encoding is performed within anatomically defined brain regions and global encoding across regions (Wang et al., 25 Aug 2025). In DUET for vision-and-language navigation, fine-scale local observation encoding is dynamically combined with coarse-scale map encoding (Chen et al., 2022). The common invariant is not the choice of backbone, but the explicit separation of short-range and long-range evidence.

3. Fusion and information exchange

The distinctive element of DLGE is not merely parallel encoding, but how the two streams are fused. The supplied literature shows several fusion families: feature-space fusion, cross-attention, hierarchical fusion, and score-level fusion.

DualGate-Net provides a particularly explicit formulation. At scale $l$, local features $U_l$, global features $S_l$, and tissue prior $P_l$ are concatenated,

$X_l = [U_l; S_l; P_l],$

and a reliability gate is estimated as

$$g_l = \sigma \left( W_{g2} * \phi \left( \mathrm{BN}(W_{g1} * X_l) \right) \right).$$

The gate modulates prior injection into both streams,

$$\hat{U}_l = U_l + g_l \odot (W_u * P_l), \qquad \hat{S}_l = S_l + g_l \odot (W_s * P_l),$$

after which the concatenated result is refined by CBAM (Tabaghsar et al., 5 Jun 2026). This formulation makes the trust placed in prior information spatially adaptive rather than fixed.

Other systems use different integration operators. DE-KAN first aligns feature dimensions and then merges global and local features by element-wise summation before KAN-based bottleneck refinement (Mustakim et al., 23 Nov 2025). SPG-CDENet uses symmetric cross-attention at corresponding encoder layers so that the global encoder attends to local features and the local encoder attends to global features, with residual connections in both directions (Tian et al., 30 Oct 2025). D-Unet uses Hierarchical Fusion between the unfixed and fixed encoders and then applies a spatial pyramid global-feature extraction module for multi-scale aggregation (Liu et al., 2020). LG-Seg concatenates the fully connected outputs of the local and global pathways and passes them through three additional fully-connected layers (Marcu et al., 2016).

Some works push fusion to the decision level. DUET computes coarse-scale and fine-scale action scores and combines them dynamically: $$\sigma_t = \operatorname{Sigmoid}\left( \operatorname{FFN}\left([\hat{v}_0; \hat{r}_0]\right) \right), \qquad s_i = \sigma_t s^c_i + (1 - \sigma_t) s^{f'}_i.$$ This makes the relative influence of global planning and local grounding step-dependent (Chen et al., 2022). In DuetGraph, the local message-passing pathway and the global attention pathway remain segregated until a learnable weighted fusion,

$$Z = \alpha \cdot Z_{\text{local}} + (1-\alpha)\cdot Z_{\text{global}},$$

which is explicitly motivated by score over-smoothing in stacked global-local reasoning (Li et al., 15 Jul 2025).

A recurrent pattern emerges: naive or static combination is repeatedly treated as inadequate. The literature instead favors gated, symmetric, hierarchical, or adaptive mechanisms that regulate when local and global evidence should dominate.

4. Priors, auxiliary supervision, and related design motifs

Many DLGE instantiations do more than split local and global encoders; they also introduce priors or auxiliary objectives to stabilize the interaction between the two streams. In DualGate-Net, tissue prior maps are integrated through a learnable prior-gated fusion mechanism rather than direct fusion, specifically because static fusion may propagate noisy information. The model also includes an auxiliary foreground reconstruction branch with an $L_1$ loss and cellness-guided cues added as an input channel, yielding a total of 6 channels: RGB + tissue prior channels + cellness (Tabaghsar et al., 5 Jun 2026).

SPG-CDENet adopts a two-stage formulation in which a spatial prior network generates coarse localization maps that delineate the approximate ROI, and the local encoder operates on the prior-guided image $U_l$0 (Tian et al., 30 Oct 2025). D-Unet similarly uses a fixed encoder to provide prior structural information, with SRM filters or DWT encoders supplying direction, edge, and multi-scale cues that assist the unfixed encoder (Liu et al., 2020). DFCRNet adds collaborative dictionary learning and multi-loss computation to balance global and local branches during mining-area scene classification (Fan et al., 27 Jul 2025).

This pattern indicates that DLGE often coexists with explicit prior modeling. A plausible implication is that the local-global split alone does not fully resolve ambiguity when one branch is noisy, weakly supervised, or semantically incomplete. Auxiliary supervision and structured priors are therefore used to constrain the fusion process.

Related but not identical mechanisms also appear in architectures that do not instantiate two full encoders. SEF-PNet’s Local-Global Context Aggregation computes global and local channel attention in parallel within each encoder block,

$U_l$1

showing a block-level rather than network-level local-global decomposition (Huang et al., 20 Jan 2025). This suggests that DLGE occupies one point in a broader design space of local-global allocation.

5. Empirical behavior across domains

Reported results in the supplied literature consistently associate dual local-global encoding with gains in robustness, discrimination, or generalization.

Paper / system	Reported result	Setting
DualGate-Net	macro F1-scores of 0.7722 on the validation set and 0.7345 on the test set	OCELOT cell detection
DE-KAN	mIoU of 94.5%, Dice coefficient of 97.1%, accuracy of 98.91%, and recall of 97.36%	2D teeth segmentation
LG-Seg	F-measure 0.9423	Massachusetts Buildings Dataset
LGENet	overall accuracy of 0.845 and average F1 score of 0.737; overall accuracy of 0.984 and average F1 score of 0.834	ISPRS and DFC2019 ALS datasets
DLGE for EEG	average macro precision, recall, and F1-score of 60.16%, 59.88%, and 59.56%	Cross-BCI-paradigm classification
DualTrack	GPE 4.93 mm, LPE 122.01 µm, FDR 5.10%, Max Drift 8.33 mm; average reconstruction error below 5 mm	Sensorless 3D ultrasound

In DualGate-Net, the reported OCELOT results are presented as evidence for adaptive prior integration in histopathology (Tabaghsar et al., 5 Jun 2026). In DE-KAN, ablation results indicate that single-encoder variants give mIoU in the 55–65% range, while the full dual-encoder model reaches 94.5% mIoU and 97.1% Dice, with up to +4.7% improvement in Dice compared to existing methods (Mustakim et al., 23 Nov 2025). In LG-Seg, the dual-stream model improves the European Buildings Dataset F-measure from 0.8266 for the local-only model and 0.6271 for the global-only model to 0.8420 for the combined model, and the architecture converges fastest (Marcu et al., 2016). In DualTrack, the full decoupled model improves over a local-only configuration whose GPE rises from 4.93 mm to 7.36 mm, and also outperforms a coupled local+global design with GPE 5.92 mm (Wilson et al., 11 Sep 2025). The EEG DLGE model shows a different empirical profile: its absolute performance is lower, but it is explicitly evaluated for processing diverse BCI paradigms without retraining and retuning (Wang et al., 25 Aug 2025).

The empirical record is therefore not uniform in magnitude, but it is uniform in direction: when the task genuinely requires both fine local evidence and wide contextual reasoning, ablations repeatedly show that removing one side of the architecture degrades performance.

6. Misconceptions, limitations, and open directions

A common simplification is to treat DLGE as synonymous with a CNN local branch plus a transformer global branch. The supplied literature does not support that restriction. DLGE-like systems include CNN+transformer pairings such as DualGate-Net, CNN+CNN pairings such as DE-KAN, learned+fixed encoder combinations such as D-Unet, Pose+RGB multimodal streams such as SEDS, region-level and whole-brain transformers in EEG DLGE, and even score-space local-global fusion in navigation and knowledge-graph reasoning (Tabaghsar et al., 5 Jun 2026, Mustakim et al., 23 Nov 2025, Liu et al., 2020, Jiang et al., 2024, Wang et al., 25 Aug 2025, Chen et al., 2022, Li et al., 15 Jul 2025).

Another misconception is that more context is automatically beneficial. Several papers frame the problem differently. DualGate-Net argues that direct fusion of tissue priors can propagate noise or uncertainty, motivating prior-gated fusion (Tabaghsar et al., 5 Jun 2026). DuetGraph argues that stacking local message passing and global attention can induce score over-smoothing, motivating segregated pathways and coarse-to-fine optimization (Li et al., 15 Jul 2025). DualTrack argues that tightly coupled local and global feature extraction restricts specialization, motivating decoupled encoders (Wilson et al., 11 Sep 2025). These works converge on a more specific claim: global information is useful only when its reliability and interaction with local evidence are controlled.

The limitations are also task-dependent. In the cross-BCI-paradigm setting, the reported macro F1-score is 59.56%, and the authors note relatively high computational demands because both local and global encoders rely on Transformer blocks (Wang et al., 25 Aug 2025). In road detection from low-quality aerial imagery, both L-Seg and LG-Seg perform far below their building-segmentation results, indicating that a local-global split does not remove data-quality bottlenecks (Marcu et al., 2016). These results caution against interpreting DLGE as a universal remedy.

The broader research direction in the supplied literature points toward adaptive local-global allocation rather than rigid dualism. DAR, for example, is not a dual-encoder architecture, but it formalizes a related asymmetry by preserving full-dimensional representations for local tokens while assigning reduced-dimensional representations to distant tokens, with perplexity within 1–2% of full-dimensional baselines at $U_l$2 and substantially better behavior than uniform reduction (Wang et al., 17 Jun 2026). This suggests that future DLGE research may continue moving toward selective capacity assignment, adaptive trust in context, and more efficient forms of global reasoning, while retaining the core principle that local detail and global structure make different demands on representation.