SPLG-Mamba: Structure-Preserving Local-Global Mamba Network for Salient Object Detection in Optical Remote Sensing Images

arXiv:2608.29626 · cs.CV · Submitted 2026-08-30 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "SPLG-Mamba: Structure-Preserving Local-Global Mamba Network for Salient Object Detection in Optical Remote Sensing Images".

Jane: Here's the gist of the paper. Salient object detection in optical remote sensing images requires dense predictions that preserve object completeness and structural continuity under complex backgrounds, scale variation, and irregular object shapes The proposed StructurePreserving Local-Global Mamba Network (SPLG-Mamba) addresses structural degradation during hierarchical feature propagation by recalibrating smoothed responses and detail residuals with SDR, assigning local modeling to shallow features and global modeling to deep features through Local-Global Mamba, and controlling cross-scale detail injection with GCSF Pruning redundant reasoning tokens SPLG-Mamba integrates Smooth-Detail Recalibration (SDR), hierarchy-aware Local-Global Mamba, and Gated Cross-Scale Fusion (GCSF) to solve the structural preservation challenge in ORSI-SOD The core idea is to manage feature propagation across the hierarchy by tailoring the modeling scope to the feature level Specifically, SDR recalibrates smoothed responses and detail residuals before state-space modeling, LocalGlobal Mamba assigns local modeling to shallow feature levels and global modeling to deep feature levels, and GCSF controls cross-scale detail injection during decoding How the training signal is built The training objective is defined as Lsup = λsLmask(S, Gs) + λbLmask(Sb, Gb) This objective balances the final saliency supervision and the auxiliary body-mask supervision The loss function Lmask combines intersection-overunion (IoU) loss

39: , Dice loss

40: , and structural similarity index measure (SSIM) loss

41: The auxiliary body output is supervised by a body mask generated from the ground-truth saliency annotation Gs by a normalized distance transform Where the gains come from Extensive experiments on ORSSD, EORSSD, and ORSI-4199 demonstrate that SPLG-Mamba achieves state-of-the-art performance and consistently improves structural completeness and continuity in challenging remote sensing scenes Quantitative comparisons show that SPLG-Mamba ranks first on five of the six metrics on ORSSD and ranks first on M, Fwβ, Fmaxβ, Emaxξ, and Fβ on EORSSD On ORSI-4199, it ranks first across all seven metrics and improves structural completeness for NSO and ISO attributes Ablation study of local and global mamba modeling The ablation study confirms the complementary roles of the SDR components, showing that removing either the smoothed response or detail residual reduces Fβ by 1.00 and 1.58 percentage points, respectively Replacing GCSF with additive top-down fusion reduces Fwβ and Fβ by 0.23 and 0.

Title and authors: Tom: So what does the actual mechanism look like? They propose that first, they use the SDR block to separate each feature into two things before they feed it into the main Mamba modeling. That means they get smoothed responses and detail residuals, and the smoothed response is supposed to give them region-level cues, while those residuals keep local structure cues for their state space modeling.

Jane: That makes sense. So if you have a really noisy texture in a shallow layer, the residual part should capture that local structure without letting it mess up the main prediction too much. But then they use LocalGlobal Mamba to assign different roles depending on whether the feature is shallow or deep.

Lu: Right. The Local Mamba blocks handle those high-resolution, shallow features by looking only at local windows, while the Global Mamba blocks model the entire deep feature map to keep that object-level consistency intact across the whole image. It’s a way of giving each layer a different kind of modeling scope.

Meng: I worry about that separation. If the SDR block fails to accurately separate those responses and residuals, you just end up with degraded predictions anyway, regardless of how smart the Mamba blocks are.

Lalam: The paper then adds Gated Cross-Scale Fusion to control how those details from different scales are injected during the final decoding step. This gate mechanism selectively combines lateral details with the top-down stream based on compatibility, which should help keep things coherent during the prediction phase.

The paper's summary: Tom: So what’s the training objective? They define it as Lsup, which balances two things: a saliency supervision loss and an auxiliary body-mask supervision loss. This auxiliary part uses a normalized distance transform from the ground truth saliency annotation to generate a body mask.

Jane: That’s smart because it doesn't just look at the final result; it enforces structural constraints during training by supervising that intermediate body mask, which helps prevent those fragmented or incomplete predictions we talked about earlier.

Lu: The loss function Lmask itself is a combination of intersection over union, Dice loss, and structural similarity index measure. It’s a multi-faceted approach to measuring how good the saliency prediction is structurally, not just pixel-wise accuracy.

Meng: When you look at the metrics they use for comparison—like MS-SSIM, Fw beta, and F beta—they claim state of the art performance on ORSSD and EORSSD. I need to know if those metrics actually mean anything when we’re talking about real-world scene understanding versus just a high score on a benchmark.

Lalam: The results are pretty strong; they rank first on five out of six metrics for ORSSD, and first on M, Fw beta, Fmax beta, Emax xi, and F beta for EORSSD. On the ORSI-four thousand one hundred ninety-nine dataset specifically, they rank first across all seven metrics and even improve structural completeness for NSO and ISO attributes.

The paper's improvements: Tom: So to wrap up on this SPLG-Mamba paper, it’s essentially about fixing the structural problems in salient object detection in optical remote sensing images by using that combination of SDR, local-global Mamba, and GCSF. It seems they’ve managed to consistently improve structural completeness and continuity even in really challenging scenes.

Jane: The implications are that for applications where you need reliable object boundaries and shapes—like monitoring infrastructure or analyzing satellite imagery—this approach could lead to much more trustworthy results than previous methods that might just find a spot but miss the shape entirely.

Lu: What I’m taking away is the way they decompose feature propagation into local and global modeling scopes, which is a robust way to handle different scales of information within a single network architecture. It shows that tailoring the scope based on feature level can solve structural degradation issues efficiently.

Meng: From an engineering view, it’s good that they show how these components work together to maintain coherence across the hierarchy. I still wonder about the computational cost of running this whole pipeline compared to simpler models.

Lalam: The authors did point out that they removed certain components in their ablation study, and removing either the smoothed response or the detail residual reduced F beta by one point or one and a half percentage points, which shows those parts are genuinely contributing to the structure preservation.

Tom: Exactly. They aren't just throwing components in; they’re showing that each part—the recalibration, the local-global split, and the fusion control—has a measurable impact on the final structural integrity of the prediction.

Jane: It’s a solid paper because it doesn't just claim high scores; it shows exactly how these mechanisms work to achieve better structure preservation when you are dealing with those irregular object shapes in remote sensing data.

Conclusion: Tom: So we're wrapping up on SPLG-Mamba, which is about using Smooth-Detail Recalibration, Local Global Mamba, and Gated Cross Scale Fusion to fix structural problems in saliency detection for remote sensing images.

Jane: Exactly. It’s tackling that big challenge of keeping the object shape complete and continuous even when the background is super complex or the objects are irregular shapes across different scales.

Lu: The core idea they push here is tailoring how the feature propagation happens depending on what level you are at—local modeling for shallow features and global modeling for deep ones. It’s a way to keep both fine detail and object consistency in the final prediction, which sounds ambitious.

Meng: From my side, I’m focused on deployment cost. If this whole structure preservation thing adds too much complexity to the inference time, it won't work in a production setting where we need real-time results. Does it run fast enough?

Lalam: I see the training objective they use balances the final saliency prediction with an auxiliary body mask loss, which means they are actively training the model to respect those structural constraints during learning. That’s a smart way to build in structure from the start.

Tom: And it seems like their experiments back this up. They showed state-of-the-art results on several metrics, like ranking first on five out of six metrics for ORSSD and EORSSD, and they even improved structural completeness for attributes like NSO and ISO on the ORSI four thousand one hundred ninety-nine dataset. That’s some solid evidence against just a high number.

Jane: It shows that these specific components—the SDR block recalibrating responses, the local-global split, and the GCSF controlling detail injection—actually make a difference in real-world scenarios, not just on paper.

Lu: The ablation study they ran was pretty telling too. Removing either the smoothed response or those detail residuals actually reduced F beta by one point or one and a half percentage points, which confirms those specific parts are doing meaningful work for structural fidelity.

Meng: But what does that mean for us when we look at the real traffic? If we try to scale this up to a massive satellite feed, do we worry about the computational load of running all that local and global modeling simultaneously?

Lalam: The authors flag that while performance is strong, they didn't explicitly detail the exact computational scaling beyond what’s implied by the model structure itself. It’s a point they left open for future work.

Tom: So to recap, SPLG-Mamba uses SDR, local-global Mamba for feature modeling scope, and GCSF to control cross scale detail injection to improve structural preservation in salient object detection.

Jane: That’s the gist of it. It’s a paper that shows how careful architectural design can directly translate into better results for hard remote sensing problems.

Lu: It gives us a strong blueprint for how we might divide the modeling tasks in future vision models—treating shallow and deep features with different mathematical scopes.

Meng: I guess for production, it means we need to balance that structural complexity against the speed of the inference pipeline before we can even consider putting it into a live system.

Lalam: It’s a strong foundation, definitely worth reading up on as we move toward more robust saliency detection in these kinds of images.

Tom: Yeah, it's definitely worth checking out for anyone working on those complex visual recognition tasks where structure matters as much as the raw pixel values.

Jane: And next time we look at a paper, keep an eye out for how they handle those internal feature relationships—that's where the real magic is happening.

Nanjing University · China Pharmaceutical University

cs.CV

Submitted: 2026-08-30

Updated: 2026-08-30

DOI: 10.1109/TGRS.2026.3749809

Code: https://github.com/yxu9910/SPLG-Mamba

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 91/100

The gist: Salient object detection in optical remote sensing images requires dense predictions that preserve object completeness and structural continuity under complex backgrounds, scale variation…

Key concepts

Smooth-Detail Recalibration (SDR)
SDR recalibrates feature responses before the main modeling step. It separates features into smoothed responses, which capture region-level cues, and detail residuals, which preserve local structural information. This prevents texture from degrading the final prediction.
Local-Global Mamba Modeling
This technique tailors the modeling scope based on feature depth. Local Mamba blocks are applied to shallow features for fine detail preservation, while Global Mamba blocks handle deep semantic features globally, ensuring both local accuracy and object consistency across the image hierarchy.
Gated Cross-Scale Fusion (GCSF)
GCSF controls how details from different scales are combined during decoding. It uses a gating mechanism to selectively inject lateral details based on compatibility with the top-down stream, allowing for controlled and meaningful cross-scale information integration.

Terminology

Summary

Salient object detection in optical remote sensing images requires dense predictions that preserve object completeness and structural continuity under complex backgrounds, scale variation, and irregular object shapes The proposed StructurePreserving Local-Global Mamba Network (SPLG-Mamba) addresses structural degradation during hierarchical feature propagation by recalibrating smoothed responses and detail residuals with SDR, assigning local modeling to shallow features and global modeling to deep features through Local-Global Mamba, and controlling cross-scale detail injection with GCSF

Pruning redundant reasoning tokens

SPLG-Mamba integrates Smooth-Detail Recalibration (SDR), hierarchy-aware Local-Global Mamba, and Gated Cross-Scale Fusion (GCSF) to solve the structural preservation challenge in ORSI-SOD The core idea is to manage feature propagation across the hierarchy by tailoring the modeling scope to the feature level Specifically, SDR recalibrates smoothed responses and detail residuals before state-space modeling, LocalGlobal Mamba assigns local modeling to shallow feature levels and global modeling to deep feature levels, and GCSF controls cross-scale detail injection during decoding

How the training signal is built

The training objective is defined as Lsup = λsLmask(S, Gs) + λbLmask(Sb, Gb) This objective balances the final saliency supervision and the auxiliary body-mask supervision The loss function Lmask combines intersection-overunion (IoU) loss [39], Dice loss [40], and structural similarity index measure (SSIM) loss [41] The auxiliary body output is supervised by a body mask generated from the ground-truth saliency annotation Gs by a normalized distance transform

Where the gains come from

Extensive experiments on ORSSD, EORSSD, and ORSI-4199 demonstrate that SPLG-Mamba achieves state-of-the-art performance and consistently improves structural completeness and continuity in challenging remote sensing scenes Quantitative comparisons show that SPLG-Mamba ranks first on five of the six metrics on ORSSD and ranks first on M, Fwβ, Fmaxβ, Emaxξ, and Fβ on EORSSD On ORSI-4199, it ranks first across all seven metrics and improves structural completeness for NSO and ISO attributes

Ablation study of local and global mamba modeling

The ablation study confirms the complementary roles of the SDR components, showing that removing either the smoothed response or detail residual reduces Fβ by 1.00 and 1.58 percentage points, respectively Replacing GCSF with additive top-down fusion reduces Fwβ and Fβ by 0. p. 923 and 0.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. StructurePreserving Local-Global Mamba Network for ORSI-SOD: This system can achieve state-of-the-art performance and consistently improve structural completeness and continuity in challenging remote sensing scenes by integrating Smooth-Detail Recalibration (SDR), hierarchy aware Local-Global Mamba, and Gated Cross Scale Fusion (GCSF).

  2. SDR Block for Feature Recalibration: The SDR block separates encoder features into smoothed responses and detail residuals before Mamba modeling, ensuring that the smoothed response provides region-level cues, while the detail residuals preserve local structure cues for state-space modeling. This prevents texture-induced background responses from degrading predictions.

  3. Local-Global Mamba Modeling: The system assigns modeling scope by using Local Mamba blocks to shallow highresolution features and Global Mamba blocks to deep semantic features, which ensures that shallow levels are modeled locally and deep levels globally, maintaining both fine detail preservation and object-level consistency across the feature hierarchy.

  4. Gated Cross Scale Fusion for Decoding: The GCSF block controls cross scale detail injection during decoding by using a gating mechanism, specifically where the gate and fused output are computed as g = SigmoidBN(Conv1×1(Q)), (1) Y = Conv3×3 g ⊙ U + (1 − g) ⊙ Te, (2), to selectively combine lateral details with the top-down stream based on compatibility.

  5. Attribute-Based Performance Enhancement: The system demonstrates stronger performance across challenging conditions by achieving improvements in metrics like MS-SSIM, Fwβ, and Fβ on attributes such as NSO (narrow salient object), ISO (incomplete salient object), and CSO (complex salient object).

Related papers