PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Yi Xu, Ruichao Hou, Tongwei Ren, Gangshan Wu

arXiv:2608.29626v1cs.CV

TL;DR

ORSI-SOD predictions can lose structural completeness during hierarchical propagation, as shallow textures, deep semantics, and cross-scale fusion affect foreground continuity. SPLG-Mamba combines feature recalibration, hierarchy-aware local-global Mamba modeling, and gated fusion, with experiments reporting state-of-the-art performance across three datasets.

  • Problem

    ORSI-SOD can suffer structural degradation when shallow details introduce background responses, deep semantics over-smooth weak structures, and cross-scale fusion disturbs coherent regions.

  • Method

    SPLG-Mamba uses SDR, hierarchy-aware Local-Global Mamba, and GCSF to recalibrate features, match modeling scope to feature levels, and control cross-scale detail injection.

  • Results

    Experiments on ORSSD, EORSSD, and ORSI-4199 demonstrate state-of-the-art performance and improved structural completeness and continuity.

  • Takeaways & Limitations

    SPLG-Mamba reports strong performance for large, complex, narrow, and incomplete objects and in cluttered or low-contrast scenes.

Abstract

from arXiv · show

Salient object detection in optical remote sensing images (ORSI-SOD) requires dense predictions that preserve object completeness and structural continuity under complex backgrounds, scale variation, and irregular object shapes. Existing methods often localize salient regions, but their predictions may still suffer from structural degradation, including fragmented, incomplete, or locally missing foreground responses. This degradation is closely related to hierarchical feature propagation, where shallow details can introduce texture-induced background responses, deep semantics may over-smooth weak structures, and uncontrolled cross-scale fusion can disturb coherent regions. To address this issue, we propose a novel Structure-Preserving Local-Global Mamba Network, SPLG-Mamba, for ORSI-SOD. Specifically, SPLG-Mamba integrates Smooth-Detail Recalibration (SDR), hierarchy-aware Local-Global Mamba, and Gated Cross-Scale Fusion (GCSF). SDR recalibrates smoothed responses and detail residuals before state-space modeling, Local-Global Mamba assigns local modeling to shallow feature levels and global modeling to deep feature levels, and GCSF controls cross-scale detail injection during decoding. Experiments on ORSSD, EORSSD, and ORSI-4199 demonstrate state-of-the-art results and improved structural completeness and continuity. The code is available at https://github.com/yxu9910/SPLG-Mamba

I. INTRODUCTION

ORSI-SOD must preserve complete and continuous salient structures despite scale, shape, contrast, and background variation. SPLG-Mamba addresses structural degradation during hierarchical propagation through recalibration, hierarchy-aware local-global modeling, and controlled cross-scale fusion.

  • ORSI-SOD supports geographic interpretation, target localization, scene understanding, and human-in-the-loop analysis.
  • Shallow features preserve local structures but respond to background textures, while deep features may suppress weak or narrow foreground parts.
  • SPLG-Mamba addresses structural degradation by combining SDR with hierarchy-aware Local-Global Mamba.
  • Local modeling is assigned to shallow detail-sensitive features and global modeling to deep semantic features.
  • SDR recalibrates smoothed responses and detail residuals before modeling, while GCSF controls cross-scale detail injection during decoding.
  • Experiments on ORSSD, EORSSD, and ORSI-4199 report state-of-the-art performance with improved structural completeness and continuity.

II. RELATED WORK

Prior ORSI-SOD research has explored boundary guidance, multiscale fusion, and frequency-aware representations. The remaining challenge is propagating these features without importing structured background noise or over-smoothing weak foreground parts.

  • Boundary and structure guidance targets complete salient regions, object contours, and foreground extents in complex ORSI scenes.
  • Multiscale fusion combines coarse semantic localization with fine spatial recovery to address objects with diverse sizes and fine structures.
  • Frequency-aware modeling separates region-level layout from edge- and texture-sensitive variations.
  • Structural preservation requires smooth-detail recalibration and controlled cross-scale propagation because local details can introduce background noise and deep semantics can over-smooth weak parts.
  • SPLG-Mamba’s framework combines a SwinV2 encoder, SDR, Local-Global Mamba, GCSF, and prediction heads.

C. Local-Global Context Modeling

Local-global context methods improve scene understanding, but prior designs often do not assign interaction scope to specific feature levels or control lateral detail injection. SPLG-Mamba organizes these choices across a four-level hierarchy.

  • Context methods use semantic interaction and long-range dependencies to distinguish salient objects from confusing backgrounds.
  • Prior local-global designs often leave feature-level interaction preferences and lateral-detail entry into the semantic stream unspecified.
  • Mamba captures long-range dependencies with linear complexity, making it suitable for spatially extensive dense prediction.
  • SPLG-Mamba uses a SwinV2 encoder, SDR, Local-Global Mamba, GCSF, and an auxiliary body output during training.
  • Local modeling is applied to shallow features, global modeling to deep features, and GCSF progressively controls lateral detail injection during decoding.

B. Smooth-Detail Recalibration

SDR separates region-level responses from local structural variations before state-space modeling. The hierarchy then uses local windows for shallow features and complete-map modeling for deep features to match their respective properties.

  • Smooth-Detail Recalibration: SDR separates each encoder feature into a smoothed response and a detail residual before Mamba modeling.
  • Smooth-Detail Recalibration: Smoothed responses provide coherent region cues, whereas detail residuals preserve boundaries, thin parts, and internal structures.
  • Smooth-Detail Recalibration: The SDR block forms these components through pooling, interpolation, subtraction, convolutional comparison, and residual channel attention.
  • Local-Global Mamba Modeling: Shallow high-resolution features use local windows because clutter can connect unrelated local responses under full-map interaction.
  • Local-Global Mamba Modeling: Deep low-resolution features use complete-map modeling because their compact semantic representations suit object-level organization.
  • Local-Global Mamba Modeling: The shared VSSBlock scans spatial sequences along multiple row-wise and column-wise directions in forward and reverse order.

D. Gated Cross-Scale Fusion

GCSF selectively injects lateral details into the top-down decoder stream to recover useful spatial structure while preserving coherent semantic regions. Its gating is paired with auxiliary body-mask supervision and a composite saliency loss during training.

  • D. Gated Cross-Scale Fusion: GCSF upsamples the top-down feature, combines it with same-level lateral features, and uses gated fusion to control detail injection.The gate is predicted from both streams, while the candidate update is derived from their cross-scale combination.
  • D. Gated Cross-Scale Fusion: Larger gate values favor candidate lateral updates, whereas smaller values preserve the upsampled semantic stream.The gate balances spatial detail against semantic guidance during decoding.
  • D. Gated Cross-Scale Fusion: Progressive fusion from Z4 through Z3, Z2, and Z1 updates fragmented regions while preserving coherent regions from unreliable shallow updates.A Global Mamba block further refines the fused feature before prediction.
  • E. Loss Function: The model jointly trains final saliency and auxiliary body predictions, with the auxiliary branch removed during inference.The body target comes from a normalized distance transform of the saliency annotation, strengthening supervision toward object interiors.
  • E. Loss Function: The training objective combines weighted final and auxiliary mask losses, with each mask loss comprising IoU, Dice, and SSIM terms.IoU and Dice supervise foreground coverage, while SSIM encourages local structural consistency.

IV. EXPERIMENTS

The experiments evaluate SPLG-Mamba on three ORSI-SOD benchmarks using complementary saliency and structural metrics under standardized implementation settings. ORSI-4199 additionally enables evaluation across nine challenging image attributes.

  • Datasets: SPLG-Mamba is evaluated on ORSSD, EORSSD, and ORSI-4199, whose training and test sets contain 600/200, 1,400/600, and 2,000/2,199 images, respectively.ORSI-4199 also provides annotations for nine attributes, including object size, contrast, complexity, multiplicity, and incompleteness.
  • Metrics: The evaluation reports mean absolute error, S-measure, weighted F-measure, and related F-measure metrics under common ORSI-SOD protocols.These metrics assess overall saliency quality from complementary perspectives.
  • Metrics: S-measure evaluates object- and region-aware structural similarity, while weighted F-measure accounts for spatial dependencies and error locations.MS-SSIM is additionally reported on ORSI-4199 to measure structural similarity across spatial scales.
  • Metrics: MS-SSIM on ORSI-4199 uses five scales with standard weights, while lower mean absolute error and higher values for the other metrics indicate better performance.The metric compares prediction and ground truth through multi-scale contrast-structure terms.
  • Implementation details: Images are resized to 384×384 for training and inference, with augmentation, ImageNet-pretrained SwinV2 initialization, and an 8 × 8 Local Mamba window.Predicted saliency maps are bilinearly resized to the original resolution before evaluation.

B. Quantitative Comparison

Across ORSSD, EORSSD, and ORSI-4199, SPLG-Mamba achieves strong quantitative performance, with results supporting structural completeness, continuity, stable threshold behavior, and efficient inference.

  • Overall comparison: SPLG-Mamba ranks first on five of six metrics on ORSSD and leads M and F wβ on EORSSD while remaining close to the best Sm.These results indicate improved region structure and foreground separation across the two datasets.
  • Threshold behavior: Precision-recall and F-measure curves remain consistently competitive across thresholds on all three datasets, indicating stable saliency responses rather than a single favorable threshold.The curves cover ORSSD, EORSSD, and ORSI-4199.
  • Complexity and speed: SPLG-Mamba reaches 61 FPS with 54.67 G FLOPs under the reported experimental settings.The decoder uses level-specific Local and Global Mamba together with lightweight GCSF; prior FPS values may differ because hardware and implementation settings vary.
  • Overall comparison: SPLG-Mamba ranks first on six of seven reported metrics on the complete ORSI-4199 test set and second on Sm.It exceeds the best competing results by 0.16, 0.90, and 0.81 percentage points on MS-SSIM, F wβ, and Fβ, respectively.
  • Attribute-based comparison: SPLG-Mamba ranks first across all seven metrics on NSO, ISO, and CSO, with gains of 1.46 and 1.26 percentage points in F wβ and Fβ on NSO.On ISO, Sm and F wβ margins reach 1.03 and 1.10 points; on CSO, Sm and F wβ gains reach 0.65 and 0.72 points.
  • Attribute-based comparison: SPLG-Mamba achieves the highest MS-SSIM on six of seven ORSI-4199 attributes, covering object scale, structure, scene complexity, and contrast.The comparison relates MS-SSIM to similarity with ground-truth maps across multiple scales.

C. Qualitative Comparison

Qualitative comparisons show that SPLG-Mamba more consistently preserves foreground coverage and structural continuity while reducing false positives across challenging ORSI-4199 scenes.

  • Qualitative comparison: Across 14 ORSI-4199 scenes, SPLG-Mamba more consistently preserves foreground coverage and elongated-structure continuity while producing fewer false positives on structured backgrounds.The comparison includes 12 representative competing methods.
  • Background interference: SPLG-Mamba retains salient regions with fewer background false positives in scenes where competing methods activate nearby structures or overextend boundaries.These errors are illustrated for small targets and structured backgrounds.
  • Foreground completeness: For broad, complex, or partially visible objects, SPLG-Mamba maintains more complete foreground regions while limiting background inclusion.Competing predictions may miss visible regions, contain internal gaps, or extend beyond object boundaries.
  • Separated regions: SPLG-Mamba preserves spatially separated salient regions more completely when competing methods omit components or activate nearby background.The examples include multiple architectural regions and separated water components.
  • Structural continuity: For elongated and irregular objects, SPLG-Mamba follows thin object paths more continuously while introducing fewer errors in surrounding background.Competing predictions omit narrow components, fragment objects, or break on adjacent linear structures.

D. Ablation Study

Ablations show that SDR, hierarchy-specific Local-Global Mamba, and GCSF each contribute to SPLG-Mamba’s performance, structural modeling, and controlled feature fusion.

  • Model variants: Removing SDR degrades performance, while retaining only smoothed responses or detail residuals reduces Fβ by 1.00 and 1.58 percentage points, respectively.The ablation supports complementary roles for region-level and local-structure cues.
  • Model variants: Replacing GCSF with additive top-down fusion reduces F wβ, supporting gated control of lateral detail injection into the top-down semantic stream.The replacement directly adds upsampled top-down and lateral features before convolutional refinement.
  • Local-Global Mamba: Removing either Local Mamba or Global Mamba lowers F-measure scores, while using one spatial scope throughout the hierarchy produces larger decreases.All-local uses windowed modeling at every level; All-global uses full-map modeling throughout.
  • Local-Global Mamba: Replacing Mamba with MHSA under the same spatial scopes reduces F wβ and Fβ by 1.27 and 2.03 percentage points, respectively.These results support both state-space modeling and hierarchy-specific local-global assignment.
  • Attribute ablation: The proposed Local-Global assignment gives the strongest results across all three metrics on BSO, NSO, ISO, and CSO, while All-local is marginally higher in Sm and MS-SSIM on SSO.The proposed assignment retains the highest F wβ on SSO.
  • Loss and backbone variants: Removing any loss term reduces performance, with the largest decrease occurring without SSIM; IoU, Dice, and body-mask supervision provide additional region or interior consistency.The complete model obtains the best or tied-best result on all six metrics across backbone variants.

E. Feature Visualization

Feature visualizations illustrate how SDR separates smooth and detail responses and how Local-Global Mamba and GCSF support coherent, structure-aware feature propagation.

  • SDR visualization: At shallow levels, SDR detail residuals retain fine boundary and thin-structure variations but also respond to repeated background textures, while smoothed responses distribute more broadly.At deeper levels, the two response types exhibit different spatial patterns for semantic feature processing.
  • Mamba visualization: Global Mamba strengthens target-aligned responses, makes activation more uniform across large objects, and maintains coherent activation across separated objects.These responses are compared with channel-projected SDR inputs.
  • Integrated behavior: The visualized design connects local structural continuity with controlled cross-scale detail injection during feature propagation.The figure set jointly examines SDR, Local-Global Mamba, and GCSF responses.

F. Failure Cases

SPLG-Mamba improves foreground completeness and structural continuity, but ambiguous boundaries and strong distractors remain challenging failure conditions.

  • Similar appearances between salient-object boundaries and adjacent man-made structures can cause slight prediction overflow around object borders.
  • In cluttered multi-object scenes, the model may preserve dominant targets while activating nearby distractors.
  • Future work targets more precise boundary discrimination and stronger separation of targets from cluttered multi-object scenes.
Loading 2608.29626v1…