G2HFNet: GeoGran-Aware Hierarchical Feature Fusion Network for Salient Object Detection in Optical Remote Sensing Images

Bin Wan; Runmin Cong; Xiaofei Zhou; Hao Fang; Chengtao Lv; Sam Kwong

cs.CVMar 13, 2026

G2HFNet: GeoGran-Aware Hierarchical Feature Fusion Network for Salient Object Detection in Optical Remote Sensing Images

Authors: Bin Wan, Runmin Cong, Xiaofei Zhou, Hao Fang, Chengtao Lv, Sam Kwong

Organizations: the School of Control Science and Engineering, Shandong University, Jinan 250061, China, and also with State Key Laboratory of Autonomous Intelligent Unmanned Systems, Shanghai 201210, China · School of Automation, Hangzhou Dianzi University, Hangzhou 310018, China · School of information engineering, Huzhou University, Huzhou 313000, China · the School of Data Science, Lingnan University, Tuen Mun, Hong Kong

Abstract

Remote sensing images captured from aerial perspectives often exhibit significant scale variations and complex backgrounds, posing challenges for salient object detection (SOD). Existing methods typically extract multi-level features at a single scale using uniform attention mechanisms, leading to suboptimal representations and incomplete detection results. To address these issues, we propose a GeoGran-Aware Hierarchical Feature Fusion Network (G2HFNet) that fully exploits geometric and granular cues in optical remote sensing images. Specifically, G2HFNet adopts Swin Transformer as the backbone to extract multi-level features and integrates three key modules: the multi-scale detail enhancement (MDE) module to handle object scale variations and enrich fine details, the dual-branch geo-gran complementary (DGC) module to jointly capture fine-grained details and positional information in mid-level features, and the deep semantic perception (DSP) module to refine high-level positional cues via self-attention. Additionally, a local-global guidance fusion (LGF) module is introduced to replace traditional convolutions for effective multi-level feature integration. Extensive experiments demonstrate that G2HFNet achieves high-quality saliency maps and significantly improves detection performance in challenging remote sensing scenarios.

Figures & tables

Fig. 1: Remote-sensing objects at different scales.

Fig. 2: The overall architecture of proposed G 2 HFNet.

Fig. 3: Architecture of the multi-scale detail enhancement module.

Fig. 4: Architecture of the dual-branch geo-gran complementary module.

Fig. 5: Architecture of the deep semantic perception module.

Fig. 6: Architecture of the local-global guidance fusion module.

Fig. 7: Visual comparison of EORSSD, ORSSD and ORSI-4199 datasets.

Dataset	EORSSD			ORSSD			ORSI-4199
Metric	$M$ $\downarrow$	$F_{\beta}$ $\uparrow$	$E_{\xi}$ $\uparrow$	$M$ $\downarrow$	$F_{\beta}$ $\uparrow$	$E_{\xi}$ $\uparrow$	$M$ $\downarrow$	$F_{\beta}$ $\uparrow$	$E_{\xi}$ $\uparrow$
PoolNet	0.0210	0.4614	0.6853	0.0359	0.6141	0.8142	0.0543	0.7353	0.8550
CSNet	0.0171	0.6284	0.8325	0.0188	0.7574	0.9049	0.0525	0.7151	0.8442
R3Net	0.0171	0.4174	0.6472	0.0401	0.7320	0.8713	0.0403	0.7769	0.8828
DPORTNet	0.0151	0.7377	0.8788	0.0222	0.7861	0.8940	0.0571	0.7514	0.8641
DNTD	0.0114	0.7192	0.8714	0.0219	0.7598	0.8862	0.0426	0.8032	0.9019

TABLE I: Quantitative comparisons with 18 methods on EORSSD, ORSSD and ORSI-4199.

Fig. 8: Illustration of the impact of MDE module (a), DGC module (b), DSP module (c) and LGF module (d).

Fig. 9: Illustration of geometric and granular branch.

Method	PoolNet	R3Net	PA-KRN	MJRBM	SUCA	ERPNet	EMFINet
Speed(FPS)	80.9	2	16	15.5	24	73.7	25
Para(M)	53.6	56.2	141.1	43.5	117.71	250.0	107.3
FLOPs(G)	97.6	47.5	617.7	101.3	56.4	227.4	480.9
Method	CorrNet	ACCoNet	HFANet	VST	MCCNet	HFCNet	Ours
Speed(FPS)	100	59.9	26	23	39.5	38	18.3
Para(M)	4.06	273.5	60.5	83.8	67.7	140.8	95.1

TABLE II: Comparison of the model complexity and the average running speed.

Settings	FLOPs(G)	Para(M)
MDE	20.5	2.3
DGC	16.3	1.4
DSP	5.3	0.012

TABLE III: FLOPs and parameter of each module.

Settings	$M$ $\downarrow$	$F_{\beta}$ $\uparrow$	$E_{\xi}$ $\uparrow$
w/o MDE	0.0059	0.8724	0.9708
w/o DGC	0.0054	0.8734	0.9782
w/o DSP	0.0047	0.8682	0.9763
w/o LGF	0.0051	0.8554	0.9701
Ours	0.0041	0.8808	0.9807

TABLE IV: Ablation study on different modules.

Settings	$M$ $\downarrow$	$F_{\beta}$ $\uparrow$	$E_{\xi}$ $\uparrow$
w/o U	0.0057	0.8669	0.9746
w/o PA	0.0058	0.8659	0.9724
Ours	0.0041	0.8808	0.9807

TABLE V: Ablation study on MDE module.

Settings	$M$ $\downarrow$	$F_{\beta}$ $\uparrow$	$E_{\xi}$ $\uparrow$
w/o Geo	0.0052	0.8781	0.9791
w/o Gran	0.0057	0.8708	0.9751
w/o GIB	0.0053	0.8721	0.9777
Ours	0.0041	0.8808	0.9807

TABLE VI: Ablation study on DGC module.

Settings	$M$ $\downarrow$	$F_{\beta}$ $\uparrow$	$E_{\xi}$ $\uparrow$
Bce	0.0099	0.6788	0.8625
Fm	0.0057	0.8902	0.9785
Iou	0.0087	0.8779	0.9765
Ours	0.0041	0.8808	0.9807

TABLE VII: Ablation study on loss function.

Settings	$M$ $\downarrow$	$F_{\beta}$ $\uparrow$	$E_{\xi}$ $\uparrow$
1	0.0052	0.8642	0.9613
1-2	0.0049	0.8713	0.9689
1-2-4	0.0044	0.8795	0.9795
Ours	0.0041	0.8808	0.9807

TABLE VIII: Ablation study on unshuffle factor.

Settings	$M$ $\downarrow$	$F_{\beta}$ $\uparrow$	$E_{\xi}$ $\uparrow$
ResNet	0.0089	0.7911	0.9576
Vgg	0.0125	0.7642	0.9015
PVT	0.0063	0.8453	0.9683
Ours	0.0041	0.8808	0.9807

TABLE IX: Ablation study on different backbone.

Explore similar work

RDNet: Region Proportion-Aware Dynamic Adaptive Salient Object Detection Network in Optical Remote Sensing Images
Mar 12, 2026Bin Wan, Runmin Cong, Xiaofei Zhou +3Salient Object Detection Multi-Scale Attention
School of Control Science and Engineering, Shandong University, Jinan 250061, China · Key Laboratory of Machine Intelligence and System Control, Ministry of Education, Jinan 250061, China · School of Automation, Hangzhou Dianzi University, Hangzhou 310018, China +2
SPLG-Mamba: Structure-Preserving Local-Global Mamba Network for Salient Object Detection in Optical Remote Sensing Images
Aug 30, 2026Yi Xu, Ruichao Hou, Tongwei Ren +1Mamba Salient Object Detection
State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, 210008, Jiangsu, China · School of Elite Biomedical Engineers and the Institute for Interdisciplinary Intelligent Pharmacy, China Pharmaceutical University, Nanjing 211198, China
ORSIFlow: Saliency-Guided Rectified Flow for Optical Remote Sensing Salient Object Detection
Mar 30, 2026Haojing Chen, Zhihang Liu, Yutong Li +3Salient Object Detection Rectified Flow
University of Electronic Science and Technology of China, Chengdu, China · Hainan University, Haikou, China · China University of Mining & Technology (Beijing), Beijing, China

Mar 12, 2026J/K move · Enter open · S save

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

Bin Wan, Runmin Cong, Xiaofei Zhou, Hao Fang+2

School of Control Science and Engineering, Shandong University, Jinan 250061, China · Key Laboratory of Machine Intelligence and System Control, Ministry of Education, Jinan 250061, China · School of Automation, Hangzhou Dianzi University, Hangzhou 310018, China+2

Salient Object Detection Multi-Scale Attention

Open paper AskPDF arXiv

Salient object detection (SOD) in remote sensing images faces significant challenges due to large variations in object sizes, the computational cost of self-attention mechanisms, and the limitations of CNN-based extractors in capturing global context and long-range dependencies. Existing methods that rely on fixed convolution kernels often struggle to adapt to diverse object scales, leading to detail loss or irrelevant feature aggregation. To address these issues, this work aims to enhance robustness to scale variations and achieve precise object localization. We propose the Region Proportion-Aware Dynamic Adaptive Salient Object Detection Network (RDNet), which replaces the CNN backbone with the SwinTransformer for global context modeling and introduces three key modules: (1) the Dynamic Adaptive Detail-aware (DAD) module, which applies varied convolution kernels guided by object region proportions; (2) the Frequency-matching Context Enhancement (FCE) module, which enriches contextual information through wavelet interactions and attention; and (3) the Region Proportion-aware Localization (RPL) module, which employs cross-attention to highlight semantic details and integrates a Proportion Guidance (PG) block to assist the DAD module. By combining these modules, RDNet achieves robustness against scale variations and accurate localization, delivering superior detection performance compared with state-of-the-art methods.

Abstract

Explore similar work

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

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

ORSIFlow: Saliency-Guided Rectified Flow for Optical Remote Sensing Salient Object Detection