Skip to content

RDNet: Region Proportion-Aware Dynamic Adaptive Salient Object Detection Network in Optical Remote Sensing Images

Conference: CVPR2026 arXiv: 2603.12215 Code: To be confirmed Area: Semantic Segmentation / Salient Object Detection Keywords: Remote sensing image salient object detection, dynamic adaptive convolution, wavelet transform, region proportion awareness, SwinTransformer

TL;DR

To address the challenge of large-scale variation in remote sensing images, this paper proposes RDNet, a region proportion-aware dynamic adaptive salient object detection network. RDNet uses a Proportion Guidance mechanism to dynamically select convolution kernel combinations of varying sizes, combined with wavelet frequency-domain interaction and a cross-attention localization module. The method achieves state-of-the-art performance across three ORSI-SOD benchmarks.

Background & Motivation

  1. Extreme scale variation in remote sensing targets: Objects in the same scene may range from very small (aircraft) to very large (stadiums). Fixed-kernel strategies cannot accommodate both cases — large kernels introduce excessive background noise, while small kernels fail to capture the full extent of large objects.
  2. High computational cost of self-attention: Existing methods perform self-attention directly on full-resolution features for inter-layer interaction, resulting in high computational complexity. Direct mixing of high- and low-frequency information also dilutes target representations.
  3. Lack of global modeling in CNN backbones: CNN-based feature extractors rely on local convolution kernels and struggle to capture global context and long-range dependencies.
  4. Uniform treatment in existing multi-scale schemes: Most methods apply identical multi-scale convolution combinations to all samples, without accounting for differences in target region proportions across images.
  5. Insufficient exploitation of mid-level contextual information: High-level features carry localization semantics and low-level features capture fine details, yet effective and lightweight interaction designs for mid-level contextual features remain lacking.
  6. Complex backgrounds in remote sensing scenes: Cluttered backgrounds and similar textures make precise boundary segmentation particularly challenging.

Method

Overall Architecture

RDNet employs SwinTransformer as the backbone to extract five-level features \(\{F_i^R\}_{i=1}^{5}\) from 384×384 inputs. Three core modules are designed to process high-, mid-, and low-level features respectively, followed by a bottom-up fusion to produce the final saliency map:

  • RPL module → High-level features \(F_4^R, F_5^R\) → Localization feature \(F^A\) + Proportion guidance \(F^G\)
  • FCE module → Mid-level features \(F_2^R, F_3^R\) → Contextual feature \(F^W\)
  • DAD module → Low-level feature \(F_1^R\) (guided by \(F^G\)) → Detail feature \(F^P\)

Region Proportion-Aware Localization Module (RPL)

  • Channel attention is applied to \(F_4^R\) and \(F_5^R\) separately (GAP + two 1×1 Conv layers + Sigmoid), with cross-multiply-add operations for channel-wise optimization.
  • Spatial attention is then applied (channel-wise Max Pool + Sigmoid), with cross-multiply-add operations for spatial-wise optimization.
  • The results are concatenated and passed through a 3×3 Conv to obtain the localization feature \(F^A\).
  • Proportion Guidance (PG) Block: \(F_5^R\) is processed via GAP → two FC layers → output \(F^G \in \mathbb{R}^{4 \times 1}\) (representing the target region proportion per batch sample), supervised by MSE Loss.

Dynamic Adaptive Detail-Aware Module (DAD)

Based on the region proportion output by PG, targets are categorized into three tiers, each with a dynamically selected convolution kernel combination:

Region Proportion Kernel Combination Design Rationale
> 50% 1×1, 3×3, 5×5, 7×7, 9×9 (5 types) Large kernels capture overall regions; small kernels refine boundaries
25%–50% 1×1, 3×3, 5×5, 7×7 (4 types) Balanced for medium-scale targets
< 25% 1×1, 3×3, 5×5 (3 types) Avoids excessive background introduction from large kernels
  • Lower branch (Detail Extractor): Multi-kernel convolutions are applied and summed → \(F_1^D\)
  • Upper branch (Detail Optimizer): Channel-wise Max Pool → same kernel combination → summed → 1×1 Conv + Sigmoid → weight \(W\)
  • Final output: \(F^P = F_1^D \otimes W \oplus F_1^D\)

Frequency-Matching Context Enhancement Module (FCE)

Wavelet Interaction Stage:

  • Discrete Wavelet Transform (DWT) is applied to \(F_2^R\) and \(F_3^R\) to obtain four frequency components (LL/LH/HL/HH).
  • Corresponding frequency components interact via matrix multiplication (reshape → transpose → matrix multiply → softmax → multiply back → IDWT), reducing computational complexity to 1/4 of full-resolution attention.

Feature Enhancement Stage:

  • Interaction results are concatenated with the original features → channel attention → spatial attention → concatenation → 3×3 Conv → \(F^W\)

Loss & Training

\[L_{total} = \frac{1}{N} \sum_{i=1}^{N} (L_{bce} + L_{iou} + L_{fm} + L_{mse})\]
  • BCE Loss: pixel-level cross-entropy
  • IoU Loss: region overlap
  • F-measure Loss: harmonic mean of precision and recall
  • MSE Loss: supervises region proportion prediction \(F^G\)

Key Experimental Results

Main Results: State-of-the-Art Across Three Benchmarks

Method EORSSD M↓ EORSSD \(F_\beta\)↑ EORSSD \(E_\xi\)↑ ORSSD M↓ ORSSD \(F_\beta\)↑ ORSSD \(E_\xi\)↑ ORSI-4199 M↓ ORSI-4199 \(F_\beta\)↑
GeleNet 0.0066 0.8367 0.9678 0.0083 0.8879 0.9787 0.0266 0.8711
ADSTNet 0.0065 0.8321 0.9633 0.0089 0.8856 0.9800 0.0319 0.8615
HFCNet 0.0051 0.7845 0.9280 0.0073 0.8581 0.9554 0.0270 0.8272
RDNet (Ours) 0.0049 0.8563 0.9718 0.0066 0.9080 0.9852 0.0254 0.8781
  • On EORSSD, MAE is reduced by 3.9% relative to HFCNet, with an average \(F_\beta\) improvement of 9.1%.
  • On ORSSD, \(F_\beta\) reaches 0.908, a 2.5% gain over ADSTNet.
  • t-test p-values against all 21 compared methods are statistically significant.

Ablation Study

Configuration M↓ \(F_\beta\)↑ \(S_\alpha\)↑
w/o DAD 0.0052 0.8550 0.9273
w/o FCE 0.0061 0.8453 0.9224
w/o RPL 0.0054 0.8561 0.9329
Full RDNet 0.0049 0.8563 0.9327
  • Removing FCE yields the largest MAE increase (0.0061 vs. 0.0049), indicating that mid-level contextual interaction contributes most.
  • Backbone comparison: SwinTransformer >> PVT > ResNet > VGG >> ViT (ViT MAE as high as 0.0175).
  • The threshold configuration [<25%, 25%–50%, >50%] is validated as optimal.

Model Efficiency

  • FLOPs: 48.7G (vs. GeleNet at 11.7G and PA-KRN at 617.7G)
  • Inference speed: 13.6 FPS (moderate speed due to intensive matrix operations)

Highlights & Insights

  1. Region proportion-guided dynamic kernel selection is the core innovation, introducing a classification paradigm into detection — predicting the target "size category" first before determining the convolution strategy, thereby avoiding kernel–scale mismatch.
  2. Wavelet-domain frequency-matching interaction transfers inter-layer feature interaction from the spatial domain to the frequency domain; interacting same-frequency components separately reduces computational cost by 4× while preventing mutual interference between high- and low-frequency information.
  3. Hierarchical three-module design (high-level localization + mid-level context + low-level detail) is logically coherent, with each module having a clearly defined role.
  4. Experiments are highly thorough: comparisons against 21 methods, 7 ablation groups, statistical significance testing via t-test, and failure case analysis.

Limitations & Future Work

  1. Slow inference speed: 13.6 FPS is insufficient for real-time remote sensing applications; intensive matrix operations are the primary bottleneck.
  2. Coarse three-tier proportion discretization: Continuous regression of region proportions may yield finer-grained adaptation than a discrete three-class scheme.
  3. PG Block relies on high-level semantics: Using only \(F_5^R\) for proportion prediction may be inaccurate for very small targets.
  4. Failure cases: Missed detections persist for extremely small or elongated targets; false detections occur when background textures are similar to targets.
  5. Validation limited to remote sensing datasets: Generalization to natural image SOD benchmarks has not been tested.
  6. Heavy SwinTransformer backbone: Limits deployment potential on edge devices.
  • vs. ADSTNet / GeleNet (prior SOTA): RDNet comprehensively outperforms both across all three datasets; the core advantage lies in the region proportion-adaptive mechanism.
  • vs. ASTT (Transformer-based method): \(F_\beta\) improves by 13.6%, benefiting from the hierarchical design rather than simple global attention.
  • vs. MCCNet / CorrNet (context interaction methods): FCE's wavelet-domain interaction is more effective and more lightweight than direct feature concatenation or attention-based interaction.
  • vs. VST (Vision Transformer): MAE is reduced by 28.9%, demonstrating that Swin's hierarchical window attention is better suited to dense prediction than ViT's flat architecture.

Rating

  • Novelty: ⭐⭐⭐⭐ — Region proportion-guided dynamic kernel selection and wavelet frequency-domain interaction are both practically meaningful innovations.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ — Comparisons against 21 methods, 7 ablation groups, statistical significance testing, and failure case analysis make this an exceptionally rigorous evaluation.
  • Writing Quality: ⭐⭐⭐⭐ — Clear structure, complete mathematical derivations, and rich figures and tables.
  • Value: ⭐⭐⭐⭐ — Provides an effective solution to the scale problem in remote sensing salient object detection, though real-time performance remains to be improved.