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

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Salient Object Detection (SOD) in optical remote sensing images (ORSI-SOD) is gaining research interest.
• Deep learning models face challenges with varying object scales, complex structures, and background interference in ORSI-SOD.

Purpose of the Study:

• To introduce a novel Cross-Model Nested Fusion Network (CMNFNet) to enhance ORSI-SOD performance.
• To leverage heterogeneous features for improved detection accuracy in complex remote sensing imagery.

Main Methods:

• The CMNFNet utilizes two heterogeneous encoders: a CNN for local features and a GCN for local and global features.
• The GCN encoder projects images into two graphs with different receptive fields for parallel graph convolutions.
• An attention-enhanced cross-model nested fusion module (AECMNFM) progressively fuses heterogeneous features and reduces background interference.

Main Results:

• The proposed CMNFNet effectively models both local and global features using its dual-encoder architecture.
• The AECMNFM adaptively refines feature representations and mitigates background noise.
• Experiments show CMNFNet outperforms 16 state-of-the-art models on benchmark datasets.

Conclusions:

• CMNFNet offers a superior approach to ORSI-SOD by effectively integrating diverse features.
• The model's ability to handle complex scenes and background interference marks a significant advancement.