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

Conference: CVPR 2026 arXiv: 2603.12215 Code: None Area: Image Segmentation Keywords: Salient Object Detection, Remote Sensing Images, Dynamic Convolutional Kernel Selection, Wavelet Transform, Region Proportion Awareness

TL;DR

This paper proposes RDNet, which employs a region proportion-aware Proportion Guidance block to estimate the area ratio of salient objects and dynamically selects combinations of 3/4/5 convolutional kernels of varying sizes for detail extraction. Combined with wavelet-domain frequency-matched context enhancement (reducing computation to 1/4) and a cross-attention localization module, RDNet comprehensively outperforms 21 state-of-the-art methods on three optical remote sensing SOD benchmarks: EORSSD, ORSSD, and ORSI-4199.

Background & Motivation

Background: Optical Remote Sensing Image Salient Object Detection (ORSI-SOD) has increasingly relied on CNN/Transformer-based multi-level feature extraction and fusion pipelines, achieving continuous performance gains on standard benchmarks.

Limitations of Prior Work:

1. Object scales in remote sensing images vary drastically (from aircraft spanning only a few pixels to stadiums occupying half the image). Existing methods apply fixed convolutional kernel combinations — large kernels introduce excessive background noise for small objects, while small kernels fail to capture the complete extent of large objects.
2. Full-resolution self-attention for cross-layer feature interaction incurs high computational cost, and directly mixing high- and low-frequency information dilutes object features.
3. CNN backbones lack the capacity for global context modeling and long-range dependency capture.

Key Challenge: Object scale is inherently uncertain, yet feature extraction strategies remain static — without knowing "how large the object is," it is impossible to select "how wide a perspective to adopt."

Goal: Adaptively select appropriate feature extraction strategies according to the proportion of the image area occupied by the salient object, while performing multi-level feature interaction efficiently.

Key Insight: Estimate the object region proportion from high-level features, then use this estimate to guide dynamic kernel selection in low-level feature extraction; perform mid-level feature interaction via frequency decomposition in the wavelet domain for dimensionality reduction.

Core Idea: Determine approximately how large the object is before deciding how to perceive it — region proportion-guided dynamic convolutional kernel selection.

Method

Overall Architecture

Given an input of \(4 \times 3 \times 384 \times 384\), a Swin Transformer backbone extracts five levels of features. High-level features \(F_4\) and \(F_5\) are fed into the RPL module for object localization and region proportion estimation; low-level feature \(F_1\) is processed by the DAD module, which dynamically selects convolutional kernels guided by the estimated proportion for detail extraction; mid-level features \(F_2\) and \(F_3\) are enhanced by the FCE module through wavelet-domain frequency-matched context enhancement. The outputs of the three modules are fused in a bottom-up manner to generate the final saliency map.

Key Designs

1. RPL (Region Proportion-aware Localization Module)

2. Function: Localizes salient objects and estimates their area proportion using high-level semantic features.

3. Mechanism: Applies alternating channel attention (GAP → two 1×1 Conv layers → Sigmoid) and spatial attention (Max Pooling → Sigmoid) on \(F_4\) and \(F_5\), then concatenates and applies a 3×3 convolution to obtain localization features.
4. PG (Proportion Guidance) block: Applies GAP followed by two fully connected layers on \(F_5\) to output a per-sample object region proportion, supervised by ground-truth proportion via MSE loss.

5. Design Motivation: Knowing "how large the object is" in advance enables the subsequent DAD module to select appropriate convolutional kernels.

6. DAD (Dynamic Adaptive Detail-perception Module)

7. Function: Dynamically selects the number and size of convolutional kernels for object detail extraction based on the region proportion output by PG.

8. Mechanism: The region proportion is discretized into three ranges — \(<25\%\) uses 3 kernel types (\(1\times1\), \(3\times3\), \(5\times5\)); \(25\%\)–\(50\%\) uses 4 types (adding \(7\times7\)); \(>50\%\) uses 5 types (adding \(9\times9\)). A dual-branch design is adopted: the lower branch performs detail extraction via summation of multi-scale convolutions, and the upper branch applies spatial attention weighting to suppress noise.

9. Design Motivation: Small objects do not require large receptive fields (which introduce background noise), while large objects require large receptive fields to capture complete regions — proportion-guided selection breaks the one-size-fits-all paradigm.

10. FCE (Frequency-matched Context Enhancement Module)

11. Function: Performs efficient cross-layer interaction among mid-level features, avoiding the high computational cost of full-resolution self-attention and the interference caused by mixing high- and low-frequency information.

12. Mechanism: DWT decomposes features into four frequency sub-bands (LL/LH/HL/HH) → attention-based interaction is performed between corresponding sub-bands → IDWT reconstructs the features → concatenated with original features → channel and spatial attention further suppress noise.
13. Design Motivation: Performing interaction in the frequency domain halves the spatial resolution in each dimension, reducing computation to 1/4 of full-resolution attention while preventing high- and low-frequency information from interfering with each other.

Loss & Training

• Total loss: \(\mathcal{L}_\text{total} = \mathcal{L}_\text{BCE} + \mathcal{L}_\text{IoU} + \mathcal{L}_{F\text{-measure}} + \mathcal{L}_\text{MSE}\), with equal weights.
• The first three terms supervise saliency map prediction (pixel-level BCE + region-level IoU + precision-recall-balanced F-measure).
• MSE supervises region proportion prediction.
• Optimizer: RMSprop; learning rate: \(1\text{e}{-5}\); batch size: 4; input resolution: \(384\times384\).

Key Experimental Results

Main Results

Dataset	Metric	RDNet	GeleNet (Prev. SOTA)	ADSTNet	HFCNet	Gain
EORSSD	MAE↓	0.0049	0.0066	0.0065	0.0051	−25.8%
EORSSD	Fβ↑	0.8563	0.8367	0.8321	0.7845	+2.3%
EORSSD	Eξ↑	0.9718	0.9678	0.9633	0.9280	+0.4%
ORSSD	MAE↓	0.0066	0.0083	0.0089	0.0073	−20.5%
ORSSD	Fβ↑	0.9080	0.8879	0.8856	0.8581	+2.3%
ORSI-4199	MAE↓	0.0254	0.0266	0.0319	0.0270	−4.5%
ORSI-4199	Fβ↑	0.8781	0.8711	0.8615	0.8272	+0.8%

RDNet achieves the best performance on all metrics across comparisons with 21 methods. All pairwise t-test p-values are \(<1\text{e}{-10}\), confirming statistical significance.

Ablation Study

Configuration	EORSSD MAE	EORSSD Fβ	Notes
Full RDNet	0.0049	0.8563	Baseline
w/o DAD	0.0052	0.8550	Dynamic kernel selection is effective
w/o FCE	0.0061	—	Largest impact; frequency-domain interaction is critical
w/o RPL	0.0054	—	Localization and proportion estimation are effective
No proportion guidance (fixed kernels)	↓	↓	Dynamic selection outperforms fixed strategy
Thresholds [25%, 50%]	Best	Best	Both wider and narrower thresholds degrade performance

Backbone comparison: Swin Transformer achieves Fβ \(0.8563 \gg\) ViT \(0.5762 \gg\) ResNet-50 \(0.7756\). The model runs at 48.7 GFLOPs and 13 FPS on an RTX 3090.

Key Findings

• The FCE module contributes the most; frequency-domain cross-layer interaction is the core driver of performance improvement.
• Region proportion-guided dynamic kernel selection consistently outperforms fixed kernel strategies.
• Swin Transformer's global context modeling capability is critical for remote sensing SOD.
• Failure cases are concentrated on extremely small objects and scenes where object texture is highly similar to the background.

Highlights & Insights

1. Region proportion → dynamic kernel selection is an intuitive and effective design — deciding "how to look" based on "how large the object is."
2. Wavelet-domain frequency-matched interaction reduces computation to 1/4 of full-resolution self-attention while preventing high- and low-frequency information from interfering with each other.
3. The PG block directly supervises proportion prediction with MSE loss, providing a well-defined learning objective for dynamic selection rather than relying on pure heuristics.
4. MAE is reduced by 4.5%–25.8% over the previous SOTA across three datasets, representing substantial improvements.

Limitations & Future Work

1. The inference speed of 13 FPS is insufficient for real-time remote sensing detection requirements.
2. The three-tier proportion thresholds (25%/50%) are manually defined; end-to-end learnable soft thresholds warrant investigation.
3. Failure cases indicate persistent difficulties with extremely small objects and objects with background-similar textures.
4. Evaluation is limited to three remote sensing SOD benchmarks without extension to natural image SOD or general segmentation tasks.

Related Work & Insights

• vs. GeleNet: Also employs Transformers for remote sensing SOD but adopts a fixed feature extraction strategy. RDNet's core advantage lies in the dynamic kernel selection mechanism that adapts to objects of varying scales.
• vs. ADSTNet: An adaptive dual-stream Transformer; Fβ of 0.8321 vs. RDNet's 0.8563, with the gap attributable to targeted handling of multi-scale objects.
• vs. HFCNet: The closest competitor in MAE (0.0051 vs. 0.0049), yet a substantial Fβ gap (0.7845 vs. 0.8563) indicates insufficient region completeness.
• Insights: The region proportion-guided paradigm could be transferred to anchor-free detectors for dynamic receptive field adjustment; wavelet-domain feature interaction could be applied to multimodal feature fusion.

Rating

• Novelty: ⭐⭐⭐⭐ Region proportion-guided dynamic kernel selection is a genuinely novel contribution, though the overall framework remains within the encoder-decoder + attention paradigm.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comparisons with 21 methods, multiple ablation groups, and t-test statistical significance verification are comprehensive.
• Writing Quality: ⭐⭐⭐ Formulas and structure are clear, though some descriptions are redundant.
• Value: ⭐⭐⭐⭐ Practically valuable within the remote sensing SOD subfield; the dynamic kernel selection idea has reasonable generalizability.

Related Papers

• [CVPR 2026] RSONet: Region-guided Selective Optimization Network for RGB-T Salient Object Detection
• [CVPR 2026] SDDF: Specificity-Driven Dynamic Focusing for Open-Vocabulary Camouflaged Object Detection
• [CVPR 2026] SGMA: Semantic-Guided Modality-Aware Segmentation for Remote Sensing with Incomplete Multimodal Data
• [CVPR 2026] FCL-COD: Weakly Supervised Camouflaged Object Detection with Frequency-aware and Contrastive Learning
• [CVPR 2026] Task-Oriented Data Synthesis and Control-Rectify Sampling for Remote Sensing Semantic Segmentation