Global Semantic-Sense Aggregation Network for Salient Object Detection in Remote Sensing Images

Area of Science:

• Computer Vision
• Geospatial Analysis
• Remote Sensing

Background:

• Salient object detection (SOD) in remote sensing images (RSI) is crucial for geographical information analysis but faces challenges like shadow interference, feature confusion, and unclear edges.
• Existing methods struggle to accurately identify salient objects due to these inherent difficulties in RSI.

Purpose of the Study:

• To develop an effective Global Semantic-aware Aggregation Network (GSANet) for improved SOD in RSI.
• To enhance the localization and semantic understanding of salient objects by addressing challenges in RSI.

Main Methods:

• Designed the Global Semantic-aware Aggregation Network (GSANet) utilizing information entropy to prioritize potential target regions.
• Proposed a Semantic Detail Embedding Module (SDEM) for adaptive fusion of multi-level features, enhancing salient region information.
• Introduced a Semantic Perception Fusion Module (SPFM) to analyze contextual and local details, improving perceptual capability and reducing semantic dilution.

Main Results:

• GSANet demonstrated outstanding performance on the ORSSD and EORSSD datasets.
• Achieved high metrics on the EORSSD dataset: 93.91% Sα, 98.36% Eξ, and 89.37% Fβ.

Conclusions:

• The proposed GSANet effectively aggregates salient information in RSI, outperforming existing methods.
• The network successfully addresses key challenges in SOD for remote sensing imagery, offering reliable support for geographical analyses.