Cross-Model Nested Fusion Network for Salient Object Detection in Optical Remote Sensing Images

Official PyTorch implementation of the TCYB 2025 paper "Cross-Model Nested Fusion Network for Salient Object Detection in Optical Remote Sensing Images".

Authors

Mingzhu Xu¹, Sen Wang¹, Yupeng Hu¹*, Haoyu Tang¹, Runmin Cong¹, Liqiang Nie²

¹ Shandong University
² Harbin Institute of Technology (Shen Zhen) * Corresponding author

Links

• Paper: IEEE Xplore
• Code Repository: GitHub

---

Table of Contents

• Introduction
• Highlights
• Method / Framework
• Project Structure
• Installation
• Checkpoints / Models
• Usage
• Saliency Maps
• Citation
• License

---

Introduction

This project is the official implementation of the paper "Cross-Model Nested Fusion Network for Salient Object Detection in Optical Remote Sensing Images".

CMNFNet proposes a novel cross-model nested fusion network that improves the performance of optical remote sensing image salient object detection (ORSI-SOD) by integrating heterogeneous features:

• Heterogeneous Encoders：Consist of a conventional CNN encoder for modeling local patterns and a specially designed Graph Convolutional Network (GCN) encoder that captures both local and global receptive fields.
• Encoder-GCN：Projects the image into two graphs with different receptive fields and performs message passing via parallel graph convolutions, enabling effective perception of objects at multiple scales and with complex topological structures.
• AECMNFM Module：A carefully designed Attention-Enhanced Cross-Model Nested Fusion Module that adaptively suppresses background interference and refines feature representations.
• Experimental Performance：Outperforms 16 state-of-the-art methods on three benchmark datasets: ORSSD, EORSSD, and ORSI-4199.

Example Description

We present CMNFNet, a framework for Salient Object Detection in Optical Remote Sensing Images.
Our method addresses complex topological structures and multi-scale perception by introducing heterogeneous encoders (CNN & GCN) and a nested fusion strategy.
This repository provides the official implementation, pretrained weights, and predicted saliency maps.

---

Highlights

• Introduces a novel heterogeneous dual-encoder architecture (CNN + GCN).
• Proposes the Encoder-GCN subnetwork to jointly model local and global context in different semantic spaces.
• Develops a Nested Fusion Module (AECMNFM) that leverages hybrid attention mechanisms to suppress background noise and fuse heterogeneous features effectively.
• Demonstrates superior performance on three challenging datasets.

---

Method / Framework

CMNFNet extracts features in parallel through two heterogeneous encoders and progressively integrates them using a nested fusion strategy.

Framework Figure

Figure 1. Overall framework of CMNFNet.

---

Project Structure

.
├── Fig/                   # Framework diagrams and visualization results
├── pretrained/            # Stores pretrained backbone weights (.pth)
├── mainNest_resnet.py     # Training/testing script with ResNet backbone
├── mainNest_vgg.py        # Training/testing script with VGG backbone
├── requirements.txt       # Environment dependencies
└── README.md

---

Installation

1. Environment

This code has been tested under PyTorch v2.0.0 + CUDA 11.8 and Python 3.8.

2. Setting Up

# 1. Create a Conda environment
conda create -n cmnfnet python=3.8
conda activate cmnfnet

# 2. Install PyTorch and graph-related libraries
pip install torch==2.0.0 torchvision==0.15.1 --index-url https://download.pytorch.org/whl/cu118
pip install torch-geometric
pip install torch-sparse torch-cluster torch-scatter -f https://data.pyg.org/whl/torch-2.0.0+cu118.html

# 3. Install other dependencies
pip install -r requirements.txt

---

Checkpoints / Models

Pretrained Weights

Please download the pretrained weights for ResNet and VGG and place them in the ./pretrained directory:

• Download Link: Baidu Drive (Password: 7kgq)

---

Usage

Training & Testing

You can run the corresponding script depending on the selected backbone:

# Using ResNet backbone
python mainNest_resnet.py

# Using VGG backbone
python mainNest_vgg.py

---

Saliency Maps

We provide saliency maps generated by CMNFNet (with both VGG and ResNet backbones) on the ORSSD, EORSSD, and ORSI-4199 datasets:

• Download Link: Baidu Drive (Password: 97ds)

---

Citation

If you use this code or method in your research, please cite our paper:

@ARTICLE{11163514,
  author={Xu, Mingzhu and Wang, Sen and Hu, Yupeng and Tang, Haoyu and Cong, Runmin and Nie, Liqiang},
  journal={IEEE Transactions on Cybernetics}, 
  title={Cross-Model Nested Fusion Network for Salient Object Detection in Optical Remote Sensing Images}, 
  year={2025},
  volume={55},
  number={11},
  pages={5332-5345},
  doi={10.1109/TCYB.2025.3571913}
}

---

License

This project is released under the Apache License 2.0.