Cross-Scale Gated Fusion (CrGF)

• Cross-scale Gated Fusion (CrGF) modules are adaptive neural architectures that integrate multi-resolution features to overcome semantic gaps and gradient instability.
• They employ multi-branch feature aggregation, channel splitting, and learnable gating to modulate the information flow and preserve detail alongside global context.
• CrGF techniques are applied in tasks like semantic segmentation, 3D detection, and reflection separation, achieving measurable gains in PSNR, mIoU, and inference speed.

Cross-scale Gated Fusion (CrGF) refers to a class of neural architectural modules that adaptively integrate information from features at multiple spatial or depth resolutions, modulating this integration with channel- or spatial-wise gates. CrGF blocks are designed to address the challenges posed by multi-scale feature aggregation, such as information degradation, semantic/information imbalance, and gradient instability in tasks ranging from semantic segmentation and monocular 3D detection to image layer separation. These mechanisms enable deep networks to preserve local detail and global context, facilitating efficient and robust learning in complex structured prediction problems.

1. Conceptual Overview and Motivation

Cross-scale Gated Fusion modules originated to solve the multi-scale coordination challenge inherent in modern deep models for vision and signal separation tasks. Multi-level feature extractors, such as fully convolutional or dual-stream networks, generate high-level (global, semantic) and low-level (local, appearance) feature maps. However, naive aggregation (summation or concatenation) introduces “semantic gaps,” causes detail loss, or leads to transmission-reflection confusion in layer separation tasks. CrGF modules overcome these deficiencies by:

• Integrating features from disparate scales or modalities via explicit gating at each spatial location or channel.
• Ensuring that only informative signals, as determined by learned gates, propagate between levels or streams.
• Enabling stable gradient propagation, selective information routing, and effective disentanglement of cross-scale representations.

Recent implementations appear across SNN-based 3D object detection (Chen et al., 9 Jun 2025), semantic segmentation (Li et al., 2019), polyp re-identification (Xiang et al., 25 Dec 2025), pneumonia detection (Wu et al., 3 Nov 2025), and image reflection separation (Lee et al., 24 Jan 2026).

2. Mathematical Formulation and Implementation Details

Representative CrGF modules utilize several core principles:

• Multi-branch feature aggregation: Inputs from different scales, depths, or sources are summed or concatenated.
• Channel splitting and bidirectional gating: Features are split into channel groups and combined multiplicatively to emphasize complementary signal flows.
• Learnable gating weights: Scalar or map-based parameters modulate each branch’s influence, typically normalized by softmax.
• Projection operations: 1×1 convolutions restore channel dimensionality or facilitate parameter sharing.
• Adaptive, data-dependent weighting: All coefficients are learned end-to-end.

The CrGF module from "ReflexSplit" (Lee et al., 24 Jan 2026) exemplifies this approach:

Let $F_{ℓ+1}$ denote decoder context from deeper level $ℓ+1$, $P_ℓ$ semantic prior from the encoder, and $E_ℓ$ texture features from another encoder branch. At decoder level $ℓ$:

• $F_\text{raw}^{ℓ} = F_{ℓ+1} + P_ℓ + E_ℓ$
• Split $F_\text{raw}^{ℓ} \rightarrow (R_1, R_2)$; split $F_{ℓ+1} \rightarrow (C_1, C_2)$ (channels halved).
• Gated branches:
  • $F_\text{main}^{ℓ} = R_1 \odot C_2$ (current-level cues gated by context),
  • $F_\text{aux}^{ℓ} = C_1 \odot R_2$ (context gated by current-level cues).
• Linear projections: $$T_1 = \varphi_1(F_\text{main}^{ℓ}), T_2 = \varphi_2(F_\text{aux}^{ℓ})$$.
• Softmax fusion: $$[\alpha, \beta] = \mathrm{softmax}( [w^{(1)}_{ℓ}, w^{(2)}_{ℓ}] )$$.
• Final fused output: $F_{ℓ}^{\text{fused}} = \alpha T_1 + \beta T_2$.

This mechanism allows adaptive routing of information relevant to both fine-grained and semantic content, stabilizing both feature encoding and gradient flow.

3. Architectural Placement and Variations

CrGF blocks are integrated at strategic locations in hierarchical architectures:

Paper	Insertion Point	Fused Sources
ReflexSplit (Lee et al., 24 Jan 2026)	Decoder at ℓ=4,3,2	Decoder context, encoder priors, texture features
SpikeSMOKE (Chen et al., 9 Jun 2025)	Between neck & head	Multi-scale spike features
GFF (Li et al., 2019)	Post-backbone	Multi-resolution backbone features
CGF-DETR (Wu et al., 3 Nov 2025)	FP neck stages	Multi-path convolutions at each FPN scale

• "ReflexSplit" applies CrGF at hierarchical decoder levels, integrating context and learned priors for layer separation.
• "SpikeSMOKE" introduces CSGC blocks after the backbone, with multi-kernel convolutions and channel attention for spike-based monocular 3D detection.
• GFF uses pixel-wise gates for fusion at each feature pyramid level to balance semantic and spatial detail (Li et al., 2019).
• The GCFC3 block (“CrGF” in (Wu et al., 3 Nov 2025)) deploys a split-transform-bypass scheme with multi-path convolutions and structural re-parameterization for fast inference.

4. Empirical Impact and Comparative Analysis

Empirical evaluations establish that CrGF variants generally outperform naive fusion mechanisms across tasks. Ablation studies from (Lee et al., 24 Jan 2026) demonstrate consistent gains in PSNR/SSIM for reflection separation:

Fusion Strategy	Real20 PSNR↑	SIR² PSNR↑	Nature PSNR↑
Direct aggreg.	24.01	25.32	25.84
Simple concat	24.89	25.67	26.21
Element-wise add	25.01	25.78	26.45
CrGF (ours)	25.22	26.33	27.03

Removing CrGF causes the network to collapse on challenging separation tasks. In semantic segmentation, fully gated fusion yields a +1.8% mean IoU improvement over naive addition or concatenation (Li et al., 2019).

CrGF blocks are analytically distinct from static concatenation (RobustSIRR), uni-directional gating (MuGI in DSRNet), and reversible-path architectures (RDNet), the last of which lacks explicit cross-scale coordination or trade-off balancing between detail and semantics (Lee et al., 24 Jan 2026). The explicit bidirectional gating and softmax weighting in CrGF endow it with superior noise suppression, detail preservation, and stability.

5. Design Considerations, Limitations, and Comparisons

The core design parameters include channel split granularity, type and number of input streams, position in the inference pipeline, gating parameter initialization, and projection operations. In "ReflexSplit," CrGF is only applied at decoder levels where semantic priors, texture details, and context co-occur. No spatial or dilated convolutions are embedded in the module itself; spatial context is encoded upstream (Lee et al., 24 Jan 2026).

Compared to attention-only or simple fusion blocks (CBAM, SENet), CrGF's channel- or spatial-wise, data-dependent multiplicative gates provide a stricter functional separation and finer modulation of information flow. A plausible implication is that CrGF reduces feature degradation and prevents signal overwhelming at shallow decoder layers, as evidenced in qualitative results (Lee et al., 24 Jan 2026).

6. Applications and Impact Across Domains

CrGF and its variants are employed in:

• Reflection separation: ReflexSplit achieves state-of-the-art transmission/reflection demixing, stabilizing gradients and resolving cross-layer confusion (Lee et al., 24 Jan 2026).
• Monocular 3D object detection: CSGC in SpikeSMOKE mitigates information loss in SNNs, boosting KITTI AP by up to +3.2 while reducing energy consumption by 72% (Chen et al., 9 Jun 2025).
• Semantic segmentation: CrGF-inspired GFF improves mIoU (+1.8) and consistently outperforms unregulated fusion on Cityscapes and ADE20K benchmarks (Li et al., 2019).
• Object detection transformers: GCFC3/CrGF enables efficient, accurate pneumonia detection with ~3.7% mAP gain and reduced latency (Wu et al., 3 Nov 2025).
• Multimodal/multilevel feature learning: While "GPF-Net" (Xiang et al., 25 Dec 2025) does not implement explicit cross-scale spatial fusion, its stacked gated modules highlight the extensibility of gating for progressive refinement.

CrGF has become foundational in tasks where multi-scale feature integration is critical to performance and robustness across diverse architectures.