RegistrationMamba: Mamba-Based Image Registration

• RegistrationMamba is a cross-modal image registration framework that uses a Mamba-based state space encoder for robust global context extraction.
• It incorporates multi-level feature aggregation and multi-expert feature learning to recover local details and enrich features under texture-limited conditions.
• The framework achieves competitive performance with linear complexity, outperforming conventional CNN and Transformer methods on optical–SAR datasets.

Searching arXiv for papers on RegistrationMamba and related Mamba-based registration methods. RegistrationMamba is a Mamba-based registration framework for cross-modal remote sensing image registration, proposed to align images from different sensors or modalities, most notably optical–SAR pairs, under pronounced nonlinear radiometric variation and limited texture (Wang et al., 6 Jul 2025). The method replaces a conventional CNN- or Transformer-dominated encoder with a visual state space encoder and combines three elements: global context feature extraction using a Mamba-based visual state space backbone, multi-level feature aggregation for local-detail recovery, and multi-expert feature learning for feature enrichment under texture-limited conditions (Wang et al., 6 Jul 2025). In the broader literature, the name also sits within an emerging class of “RegistrationMamba”-style systems in which state space models serve as the registration backbone for dense medical registration, multi-modal deformable alignment, or point cloud registration, typically to obtain long-range dependency modeling with linear complexity rather than quadratic self-attention (Guo et al., 2024, Wang et al., 2024, Wen et al., 2024, Liu et al., 16 Jun 2025).

1. Definition and problem setting

RegistrationMamba is designed for cross-modal remote sensing image registration, a task in which images from distinct sensing modalities must be aligned despite modality-dependent appearance differences and limited discriminative texture (Wang et al., 6 Jul 2025). The paper emphasizes two central difficulties. The first is nonlinear radiometric variation, described as a modality gap arising because optical and SAR images are generated by fundamentally different imaging physics. The second is limited texture and repetitive local patterns, which makes discriminative correspondence extraction difficult, especially in homogeneous or high-resolution regions (Wang et al., 6 Jul 2025).

The framework is positioned against two dominant design families. CNN-based registration methods are described as being limited by the local receptive field and therefore weak in long-range contextual modeling. Transformer-based methods improve global modeling but incur quadratic self-attention cost, which restricts their applicability to high-resolution cross-modal remote sensing imagery (Wang et al., 6 Jul 2025). RegistrationMamba is presented as a middle-ground design in which global modeling is achieved through state space sequence processing with linear complexity, while local discriminability is strengthened by dedicated aggregation and expert-fusion modules (Wang et al., 6 Jul 2025).

A broader reading of the literature suggests that RegistrationMamba is both a specific method and an instance of a wider research direction: using Mamba or related state space models as the global registration backbone in settings where registration depends on long-range correspondence but cannot tolerate the cost of full attention (Guo et al., 2024, Wang et al., 2024, Liu et al., 16 Jun 2025). This suggests that the term can denote not only the specific optical–SAR architecture, but also a broader architectural motif in registration research.

2. Architectural composition

The framework has three major parts: a Mamba-based visual state space encoder for global context feature extraction, a multi-level feature aggregation module for local texture refinement, and a multi-expert feature learning strategy for richer representations under texture-limited conditions (Wang et al., 6 Jul 2025). The encoder uses a 2D selective scanning module over image features, while the aggregation and expert-learning components address the local-detail and representation-enrichment deficiencies that would remain if one relied on global sequence modeling alone (Wang et al., 6 Jul 2025).

The encoder is hierarchical. Input images are first processed by a $7\times7$ convolutional block, then patchified and passed through multiple visual state space blocks, with downsampling between stages (Wang et al., 6 Jul 2025). The paper specifies three feature scales, with two cascaded VSS blocks per scale and output channels $96$, $192$, and $384$ (Wang et al., 6 Jul 2025). This design couples large receptive-field propagation with a multi-scale feature hierarchy.

The multi-expert feature learning component is explicitly motivated by texture scarcity. An image transformation pool generates augmented image variants through flipping, random rotation, and homography transformation, producing transformed inputs $X_S^n=T_n(X_S)$ (Wang et al., 6 Jul 2025). Each transformed image is passed through RegistrationMamba to obtain features $F_S^n$, which are then processed by expert-specific modules $E_n$ and fused by a learnable soft router (Wang et al., 6 Jul 2025). Unlike sparse mixture-of-experts designs that select only a subset of experts, the paper states that MEFL fuses information from all experts for richer feature representation (Wang et al., 6 Jul 2025).

The multi-level feature aggregation module integrates global contextual features with local texture details. It consists of MSAA, used at most encoder levels, and CA, used only for the deepest feature level $F'_3$ (Wang et al., 6 Jul 2025). MSAA uses multi-scale depthwise convolutions and adaptive gating, with channel split ratios

$$\delta_1=\frac{3}{8},\quad \delta_2=\frac{1}{2},\quad \delta_3=\frac{1}{8},$$

while CA performs lightweight channel reallocation at the most semantically abstract level (Wang et al., 6 Jul 2025). After MFA, multi-level features are upsampled, concatenated, and adaptively fused to produce final template and reference representations (Wang et al., 6 Jul 2025).

3. State space formulation and cross-scanning mechanism

RegistrationMamba adopts the standard continuous-time state space formulation

$$\left\{\begin{array}{l} h^{\prime}(t)=\mathbf{A} h(t)+\mathbf{B} x(t), \ y(t)=\mathbf{C} h(t)+\mathbf{D} x(t), \end{array}\right.$$

followed by a Mamba-style discretization using zero-order hold (Wang et al., 6 Jul 2025). The discrete form is given as

$96$0

with the state update

$96$1

The sequence can also be written as a convolution with an effective kernel $96$2 (Wang et al., 6 Jul 2025).

The critical Mamba-specific distinction is selective scanning, in which parameters become input-dependent: $96$3 The paper describes this as converting the model into a linear time-varying system so that state updates adapt to the input content (Wang et al., 6 Jul 2025).

For 2D imagery, RegistrationMamba uses a 2D selective scanning module, SS2D. Given input feature $96$4, the formulation is

$96$5

where $96$6 forms sequences from the 2D feature map in four scanning directions, $96$7 denotes the selective cross-scanning state sequential model, and $96$8 recombines the outputs into a 2D feature map (Wang et al., 6 Jul 2025). This multi-directional cross-scanning is the mechanism by which the framework captures image-wide contextual relationships with linear complexity.

This design places RegistrationMamba within a larger class of visual state space adaptations. VMambaMorph extends a visual state space cross-scan block to 3D volumetric registration, aiming to propagate information across axial, sagittal, and coronal dependencies in brain MR–CT registration (Wang et al., 2024). MT-PCR uses Z-order serialization to convert unordered 3D point clouds into sequences that Mamba can process for coarse matching before Transformer refinement (Liu et al., 16 Jun 2025). These related systems indicate that the central challenge is not only replacing attention, but adapting state space processing to the geometry of the registration domain.

4. Feature enrichment and matching objective

MEFL is one of the distinctive components of RegistrationMamba. For each transformed template image $96$9, the framework extracts features $192$0, then passes them through expert modules: $192$1 The expert outputs are fused through a learnable soft router, with router parameters $192$2 mapped to weights

$192$3

and final aggregated feature

$192$4

The paper states that soft fusion outperforms both sparse top-2 expert selection and equal weighting, and that increasing the number of experts from $192$5 to $192$6 improves performance, with $192$7 experts chosen as the best trade-off (Wang et al., 6 Jul 2025).

For similarity computation, the template feature map is treated as a convolution kernel and slid over the reference feature map: $192$8 The normalized similarity is then

$192$9

The paper states that this convolutional correlation is GPU-friendly and faster than pixel-wise matching (Wang et al., 6 Jul 2025).

The training objective follows F3Net and consists of three losses. The matching loss uses a $384$0 positive region around the ground-truth location and the top-$384$1 hardest negatives with $384$2: $384$3 The fine similarity loss uses a Gaussian target and top-$384$4 supervision with $384$5: $384$6 and the similarity peak loss is

$384$7

The final objective is

$384$8

with $384$9 and $X_S^n=T_n(X_S)$0 (Wang et al., 6 Jul 2025).

A plausible implication is that RegistrationMamba does not treat long-range modeling, local detail, and matching supervision as interchangeable. Instead, global context is handled by the VSS encoder, local texture by MFA, and ambiguity reduction by MEFL and a correlation-based similarity objective.

5. Datasets, implementation, and empirical results

The method is evaluated on two optical–SAR datasets with different resolutions (Wang et al., 6 Jul 2025). SEN1-2 contains $X_S^n=T_n(X_S)$1 co-registered optical/SAR patch pairs from Sentinel-1 and Sentinel-2 at $X_S^n=T_n(X_S)$2 resolution and image size $X_S^n=T_n(X_S)$3. Following prior work, $X_S^n=T_n(X_S)$4 pairs per season are sampled, yielding $X_S^n=T_n(X_S)$5 total pairs, which are split $X_S^n=T_n(X_S)$6 into $X_S^n=T_n(X_S)$7 training and $X_S^n=T_n(X_S)$8 testing samples. The SAR template crop size is $X_S^n=T_n(X_S)$9 (Wang et al., 6 Jul 2025). The OS dataset includes GF-3 SAR and Google Earth optical imagery at $F_S^n$0 resolution, with OS-512 containing $F_S^n$1 pairs of size $F_S^n$2 and OS-256 containing $F_S^n$3 pairs of size $F_S^n$4; OS-512 is split into $F_S^n$5 train and $F_S^n$6 test samples with template crop $F_S^n$7, and OS-256 into $F_S^n$8 train and $F_S^n$9 test samples with template crop $E_n$0 (Wang et al., 6 Jul 2025).

Evaluation uses the $E_n$1 distance between predicted and ground-truth matching locations,

$E_n$2

and Correct Matching Rate,

$E_n$3

where a match is correct if $E_n$4 (Wang et al., 6 Jul 2025). Implementation uses PyTorch on an NVIDIA RTX 4090 with AdamW, batch size $E_n$5, initial learning rate $E_n$6, $E_n$7 training epochs, $E_n$8 experts in MEFL, three scales, two VSS blocks per scale, output channels $E_n$9, and loss weights $F'_3$0 (Wang et al., 6 Jul 2025).

The principal reported results are summarized below.

Benchmark	Key results	Additional notes
SEN1-2	$F'_3$1, CMR(1) $F'_3$2, CMR(2) $F'_3$3, CMR(3) $F'_3$4, CMR(5) $F'_3$5	Best $F'_3$6 among compared methods (Wang et al., 6 Jul 2025)
OS-512	$F'_3$7, CMR(1) $F'_3$8, CMR(2) $F'_3$9, CMR(3) $$\delta_1=\frac{3}{8},\quad \delta_2=\frac{1}{2},\quad \delta_3=\frac{1}{8},$$0, CMR(5) $$\delta_1=\frac{3}{8},\quad \delta_2=\frac{1}{2},\quad \delta_3=\frac{1}{8},$$1	Outperforms F3Net and DC-InfoNCE, especially at stricter thresholds (Wang et al., 6 Jul 2025)
OS-256	$$\delta_1=\frac{3}{8},\quad \delta_2=\frac{1}{2},\quad \delta_3=\frac{1}{8},$$2, CMR(1) $$\delta_1=\frac{3}{8},\quad \delta_2=\frac{1}{2},\quad \delta_3=\frac{1}{8},$$3, CMR(2) $$\delta_1=\frac{3}{8},\quad \delta_2=\frac{1}{2},\quad \delta_3=\frac{1}{8},$$4, CMR(3) $$\delta_1=\frac{3}{8},\quad \delta_2=\frac{1}{2},\quad \delta_3=\frac{1}{8},$$5, CMR(4) $$\delta_1=\frac{3}{8},\quad \delta_2=\frac{1}{2},\quad \delta_3=\frac{1}{8},$$6, CMR(5) $$\delta_1=\frac{3}{8},\quad \delta_2=\frac{1}{2},\quad \delta_3=\frac{1}{8},$$7	$$\delta_1=\frac{3}{8},\quad \delta_2=\frac{1}{2},\quad \delta_3=\frac{1}{8},$$8M parameters, $$\delta_1=\frac{3}{8},\quad \delta_2=\frac{1}{2},\quad \delta_3=\frac{1}{8},$$9 GFLOPs, $$\left\{\begin{array}{l} h^{\prime}(t)=\mathbf{A} h(t)+\mathbf{B} x(t), \ y(t)=\mathbf{C} h(t)+\mathbf{D} x(t), \end{array}\right.$$0 ms average time per pair (Wang et al., 6 Jul 2025)

On SEN1-2, the paper reports improvement in CMR(3) from $$\left\{\begin{array}{l} h^{\prime}(t)=\mathbf{A} h(t)+\mathbf{B} x(t), \ y(t)=\mathbf{C} h(t)+\mathbf{D} x(t), \end{array}\right.$$1 for MARU-Net to $$\left\{\begin{array}{l} h^{\prime}(t)=\mathbf{A} h(t)+\mathbf{B} x(t), \ y(t)=\mathbf{C} h(t)+\mathbf{D} x(t), \end{array}\right.$$2, from $$\left\{\begin{array}{l} h^{\prime}(t)=\mathbf{A} h(t)+\mathbf{B} x(t), \ y(t)=\mathbf{C} h(t)+\mathbf{D} x(t), \end{array}\right.$$3 for F3Net to $$\left\{\begin{array}{l} h^{\prime}(t)=\mathbf{A} h(t)+\mathbf{B} x(t), \ y(t)=\mathbf{C} h(t)+\mathbf{D} x(t), \end{array}\right.$$4, and a $$\left\{\begin{array}{l} h^{\prime}(t)=\mathbf{A} h(t)+\mathbf{B} x(t), \ y(t)=\mathbf{C} h(t)+\mathbf{D} x(t), \end{array}\right.$$5-point CMR(3) improvement over FFT U-Net (Wang et al., 6 Jul 2025). On OS-256, a framework comparison states that UNet is lighter but weaker, while TransUNet and U-Transformer benefit from global modeling but are far more expensive; RegistrationMamba is reported to achieve better accuracy than these baselines while being much more efficient than Transformer-based alternatives (Wang et al., 6 Jul 2025).

The paper also reports average registration time on SEN1-2: NCC $$\left\{\begin{array}{l} h^{\prime}(t)=\mathbf{A} h(t)+\mathbf{B} x(t), \ y(t)=\mathbf{C} h(t)+\mathbf{D} x(t), \end{array}\right.$$6 ms, MI $$\left\{\begin{array}{l} h^{\prime}(t)=\mathbf{A} h(t)+\mathbf{B} x(t), \ y(t)=\mathbf{C} h(t)+\mathbf{D} x(t), \end{array}\right.$$7 ms, DDFN $$\left\{\begin{array}{l} h^{\prime}(t)=\mathbf{A} h(t)+\mathbf{B} x(t), \ y(t)=\mathbf{C} h(t)+\mathbf{D} x(t), \end{array}\right.$$8 ms, Siamese CNN $$\left\{\begin{array}{l} h^{\prime}(t)=\mathbf{A} h(t)+\mathbf{B} x(t), \ y(t)=\mathbf{C} h(t)+\mathbf{D} x(t), \end{array}\right.$$9 ms, FFT U-Net $96$00 ms, MARU-Net $96$01 ms, and RegistrationMamba $96$02 ms (Wang et al., 6 Jul 2025). Under Gaussian noise added to OS optical images, the method is reported to remain stable from $96$03 to $96$04 variance and to degrade only slightly beyond $96$05 variance (Wang et al., 6 Jul 2025). It is additionally reported to be less sensitive to template-size reduction than F3Net (Wang et al., 6 Jul 2025).

6. Ablation evidence and relation to the wider Mamba registration literature

The ablation study on OS-256 isolates the effect of the main modules (Wang et al., 6 Jul 2025). Starting from a base CNN U-Net with $96$06 and CMR(3) $96$07, adding Mamba yields $96$08 and CMR(3) $96$09, adding MEFL yields $96$10 and CMR(3) $96$11, and the full model reaches $96$12 and CMR(3) $96$13 (Wang et al., 6 Jul 2025). MFA ablations show that removing MSAA gives $96$14 and CMR(3) $96$15, removing CA gives $96$16 and CMR(3) $96$17, removing both gives $96$18 and CMR(3) $96$19, and full MFA restores the best $96$20 and CMR(3) $96$21 (Wang et al., 6 Jul 2025). Expert-number ablations show $96$22 experts at $96$23, CMR(3) $96$24; $96$25 experts at $96$26, CMR(3) $96$27; $96$28 experts at $96$29, CMR(3) $96$30; and $96$31 experts as the best setting with $96$32, CMR(3) $96$33, after which performance saturates (Wang et al., 6 Jul 2025).

These results support the paper’s claim that the framework succeeds through the joint action of linear-time global context modeling, local texture enhancement, and multi-view expert feature enrichment (Wang et al., 6 Jul 2025). They also clarify a recurring theme in Mamba-based registration research: Mamba alone is typically not the whole method. In MambaMorph, a simple fine-grained U-Net extractor is paired with a Mamba-based registration module, and the combination is reported to improve SR-Reg Dice from $96$34 for $96$35 to $96$36 for the full model (Guo et al., 2024). In VMambaMorph, a hybrid VMamba-CNN backbone plus recursive refinement yields Dice $96$37 and HD$96$38 $96$39 on SR-Reg, surpassing the reported MambaMorph baseline on that benchmark (Wang et al., 2024). In MambaReg, disentanglement of modality-dependent and modality-invariant features is combined with Bi-Mamba and ROI-aware training, producing Dice $96$40, MSE $96$41, NCC $96$42, and SSIM $96$43 on the reconstructed MSU-PID RGB-IR dataset (Wen et al., 2024). In MT-PCR, Z-order serialization and an optimized Mamba encoder are used for scalable coarse matching, with Transformer refinement retained for sparse and dense point cloud registration (Liu et al., 16 Jun 2025).

A common misconception would be that “RegistrationMamba” simply means substituting a Transformer with Mamba. The literature summarized here indicates a more specific pattern. In each case, the state space component is adapted to the geometry of the registration domain and paired with modules that address domain-specific deficiencies: feature extraction for multi-modal medical volumes (Guo et al., 2024), cross-scan redesign and recursion for 3D volumetric alignment (Wang et al., 2024), disentangled sparse coding and ROI emphasis for RGB–IR plant images (Wen et al., 2024), or spatial serialization plus Transformer refinement for point clouds (Liu et al., 16 Jun 2025). RegistrationMamba follows the same pattern through SS2D scanning, MFA, and MEFL (Wang et al., 6 Jul 2025).

7. Significance, limitations, and research context

RegistrationMamba is significant within cross-modal remote sensing because it addresses both of the paper’s stated failure modes—modality gap and texture scarcity—within a single linear-complexity framework (Wang et al., 6 Jul 2025). The visual state space encoder supplies global dependency modeling, MFA recovers local discriminability, and MEFL enriches features by aggregating transformed image variants through multiple experts (Wang et al., 6 Jul 2025). This architecture is explicitly intended to preserve the computational advantages of Mamba over self-attention while avoiding the locality limitations of pure CNN designs (Wang et al., 6 Jul 2025).

The paper’s stated limitation is that performance still requires improvement in extremely sparse-texture and highly repetitive-texture scenes (Wang et al., 6 Jul 2025). This is important because those cases are structurally aligned with the motivating difficulties of CRSI registration itself. A plausible implication is that, although global sequence modeling reduces ambiguity, it does not entirely eliminate the need for stronger inductive bias or richer supervision when local evidence is intrinsically weak or repetitive.

Within the wider literature, RegistrationMamba exemplifies an ongoing shift from attention-centric registration backbones toward state space alternatives that aim to retain global reasoning while improving scaling behavior. MambaMorph frames this in terms of deformable MR–CT registration and efficient long-range spatial correspondence modeling in 3D volumes (Guo et al., 2024). VMambaMorph extends the idea to visual state space cross-scan modules and recursive refinement (Wang et al., 2024). MambaReg emphasizes interpretability through disentanglement of modality-independent and modality-dependent features (Wen et al., 2024). MT-PCR shows that when the underlying data are unordered, a registration-capable Mamba architecture may require explicit geometric serialization before selective scanning becomes effective (Liu et al., 16 Jun 2025). RegistrationMamba contributes the remote-sensing counterpart of this trajectory: a cross-modal, optical–SAR registration system centered on multi-directional selective scanning, multi-level local aggregation, and multi-expert augmentation-aware feature fusion (Wang et al., 6 Jul 2025).