Dual-Scale Fusion Mamba Block (DFMB)

• The paper introduces DFMB, a module that fuses original high-resolution and bilinearly downsampled features via position-aligned grouping to capture both global context and local details.
• DFMB processes dual scales in parallel through dedicated paths, ensuring that structural details and broader contextual information are effectively integrated.
• Empirical results demonstrate that DFMB enhances shadow removal performance with improved PSNR (38.52) and reduced RMSE (4.37), leading to smoother boundaries and fewer artifacts.

to=arxiv_search 大发快三开奖结果 大发快三计划{"query":"arXiv (Li et al., 18 Aug 2025) D2-Mamba Dual-Scale Fusion and Dual-Path Scanning with SSMs for Shadow Removal", "max_results": 5, "sort_by": "relevance"} to=arxiv_search 大发官网 菲龙թաց 天天中彩票如何assistant to=arxiv_search code 天天中彩票不能{"query":"(Li et al., 18 Aug 2025)", "max_results": 10, "sort_by": "relevance"} to=arxiv_search 大发快三走势图{"query":"(Li et al., 18 Aug 2025)", "max_results": 3} In "D2-Mamba: Dual-Scale Fusion and Dual-Path Scanning with SSMs for Shadow Removal," the Dual-Scale Fusion Mamba Block (DFMB) is a specially designed module for multi-scale feature fusion within a Mamba-based shadow removal network (Li et al., 18 Aug 2025). It is introduced for a restoration setting in which degradation is spatially localized and non-uniform, so the transformation required for shadowed areas can differ significantly from that of well-lit regions. Within this formulation, DFMB fuses original high-resolution features and downsampled low-resolution features in a position-aligned, cross-scale manner, with the stated aims of enhancing multi-scale feature representation, propagating contextual information, and effectively reducing boundary artifacts (Li et al., 18 Aug 2025).

1. Definition and architectural setting

DFMB is motivated by the observation that shadow removal requires both global semantics and local detail modeling. The provided formulation states that prior methods either insufficiently fused features of different scales or lacked alignment between scales. DFMB addresses these limitations by fusing original and downsampled features in a position-aligned, cross-scale manner, enabling both global context propagation and precise boundary detail recovery (Li et al., 18 Aug 2025).

The block is situated after the encoder, before the main UNet architecture built with Dual-Path Mamba Group (DPMG) blocks. Its input is the feature map extracted by the encoder from the concatenation of the shadow image and binary mask, denoted

$$F_{\text{input}} \in \mathbb{R}^{C \times H \times W}.$$

Its output is a fused multi-scale feature representation, denoted $F_{\text{fused}}$, which is passed to downstream UNet and DPMG stages (Li et al., 18 Aug 2025).

2. Input construction and dual-scale processing

The input feature extraction is given as

$$F_{\text{input}} = \text{Flatten}\left(\text{LeakyReLU}(\text{Conv2d}(\text{Concat}(I_s, M)))\right).$$

A downsampled feature map is then obtained by bilinear interpolation at half resolution:

$$F_{\text{down}} = \text{BilinearInterp}(F_{\text{input}}, \text{scale}=0.5).$$

These two scales are processed in parallel, and the description specifies that both scales pass independently through a DPMG, with each path capturing global and region-level dependencies (Li et al., 18 Aug 2025).

This two-stream setup is central to the block’s formulation. The original feature carries high-resolution local structure, while the downsampled feature supplies lower-resolution contextual information. The stated rationale is that this combination enables both global context propagation and precise boundary detail recovery, which is particularly relevant when shadow boundaries and textures require transformations that differ from those of non-shadow regions (Li et al., 18 Aug 2025).

3. Position-aligned cross-scale fusion

The core DFMB mechanism is position-aligned unfolding and grouping. For a pixel $x_d(i,j)$ in the downsampled feature map, DFMB gathers the four spatially aligned neighbors from the original-resolution feature map:

$x(2i,2j),\ x(2i+1,2j),\ x(2i,2j+1),\ x(2i+1,2j+1).$

These four high-resolution features and the corresponding low-resolution feature are stacked into a five-element token group:

$$S_{i,j} = [x(2i,2j),\ x(2i+1,2j),\ x(2i,2j+1),\ x(2i+1,2j+1),\ x_d(i,j)].$$

The collection of $S_{i,j}$ units is then row-wise unfolded into a one-dimensional sequence $S$ (Li et al., 18 Aug 2025).

The sequence is processed by a Mamba block:

$F_{\text{fused}} = \text{Mamba}(S).$

The description identifies this as the key innovation of DFMB. Traditional schemes are characterized as approaches that may just add or concatenate multi-resolution features or rely on convolutions with limited spatial correspondence, whereas DFMB’s grouping ensures that each low-resolution feature is fused with its corresponding high-resolution patch. The stated effect is preservation of spatial context, minimization of misalignment artifacts, and an output feature map that is both globally aware and locally precise (Li et al., 18 Aug 2025).

4. Functional role in shadow removal

The role of DFMB inside D2-Mamba is to provide the network with strong, gradient-stable, context-rich multiscale features before the UNet backbone. In the subsequent upsampling path, skip connections further leverage these fused outputs for detailed shadow region recovery (Li et al., 18 Aug 2025).

The broader network-level context is that D2-Mamba combines dual-scale fusion with dual-path scanning. The abstract describes the network as one that selectively propagates contextual information based on transformation similarity across regions. In this larger design, DFMB is responsible for enhanced multi-scale feature representation, while the DPMG captures global features via horizontal scanning and incorporates a mask-aware adaptive scanning strategy to improve structural continuity and fine-grained region modeling (Li et al., 18 Aug 2025).

A common simplification is to treat DFMB as merely a generic multi-scale fusion layer. The provided description is more specific. DFMB is not defined simply by downsampling and concatenation; it is defined by position-aligned grouping of a low-resolution token with its corresponding $F_{\text{fused}}$0 neighborhood in the original-resolution feature map, followed by Mamba-based sequential interaction. That formulation is directly tied to the stated objective of reducing boundary artifacts in shadow removal (Li et al., 18 Aug 2025).

5. Empirical evidence and reported effects

The paper reports that the overall method significantly outperforms existing state-of-the-art approaches on shadow removal benchmarks, and the details attribute part of this result to DFMB’s multi-scale fusion behavior (Li et al., 18 Aug 2025). The ablation study is summarized as showing that removing DFMB causes a significant drop in PSNR and an increase in RMSE, especially in regions with fine shadow details. The full model is reported with PSNR (S) = 38.52 and RMSE (S) = 4.37, while a reduced variant is reported at 38.17/4.48; the summary states that the full model gives maximum performance (Li et al., 18 Aug 2025).

The qualitative evidence is described in terms of smoother transitions at shadow boundaries and fewer visual artifacts. Figures 3 and 4 are stated to show that images processed with DFMB yield cleaner and more natural shadow-free results than state-of-the-art approaches. The mechanism offered for this behavior is that cross-scale fusion with position alignment provides consistent context on both sides of shadow boundaries, while sequential interaction in the Mamba block smooths distribution mismatches at shadow edges (Li et al., 18 Aug 2025).

6. Terminological scope and related Mamba fusion blocks

The acronym "DFMB" is not used uniformly across the Mamba literature. In the source material, it denotes different modules in different papers, and several related works pursue dual-scale, dual-phase, cross-modal, or multi-domain fusion with different block names. This makes terminological disambiguation necessary when reading recent arXiv papers.

Paper	Module name	Core role
"D2-Mamba: Dual-Scale Fusion and Dual-Path Scanning with SSMs for Shadow Removal" (Li et al., 18 Aug 2025)	Dual-Scale Fusion Mamba Block (DFMB)	Position-aligned fusion of original and downsampled features for shadow removal
"Spatial-Frequency Enhanced Mamba for Multi-Modal Image Fusion" (Sun et al., 10 Nov 2025)	Dynamic Fusion Mamba Block (DFMB)	Dynamic integration of features from different branches
"CloudMamba: An Uncertainty-Guided Dual-Scale Mamba Network for Cloud Detection in Remote Sensing Imagery" (Yang et al., 8 Apr 2026)	Dual-Scale Mamba (DS-Mamba) Block	Small-scale and large-scale branches, concatenation, SS2D, residual fusion
"FusionMamba: Efficient Remote Sensing Image Fusion with State Space Model" (Peng et al., 2024)	FusionMamba block	Dual-input fusion of spatial and spectral features

Related papers further broaden the design space. "MambaDFuse: A Mamba-based Dual-phase Model for Multi-modality Image Fusion" presents a dual-phase feature fusion module consisting of shallow fusion with channel exchange and deep fusion with enhanced Multi-modal Mamba (M3) blocks (Li et al., 2024). "MSFMamba: Multi-Scale Feature Fusion State Space Model for Multi-Source Remote Sensing Image Classification" introduces a Fus-Mamba block that extends the original Mamba architecture to accommodate dual inputs and enhance cross-modal feature interactions (Gao et al., 2024). "Fusion-Mamba for Cross-modality Object Detection" uses a Fusion-Mamba block containing State Space Channel Swapping and Dual State Space Fusion for shallow and deep fusion in a hidden state space (Dong et al., 2024). By contrast, "Dual-Domain Homogeneous Fusion with Cross-Modal Mamba and Progressive Decoder for 3D Object Detection" explicitly states that the specific term "Dual-Scale Fusion Mamba Block" does not appear in the paper; its dual-scale or dual-domain fusion is instead realized through HVF and HBF with intra-modal and cross-modal Mamba blocks (Hu et al., 12 Mar 2025).

Within this broader landscape, the D2-Mamba DFMB is distinguished by its task-specific, position-aligned cross-scale tokenization strategy. The supplied materials therefore support a narrow technical meaning for DFMB in shadow removal, while also showing that the same acronym can denote different fusion mechanisms in adjacent Mamba-based literatures.