Salient Object Detection in Optical Remote Sensing Images Driven by Transformer.

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Existing Salient Object Detection in Optical Remote Sensing Images (ORSI-SOD) methods primarily use Convolutional Neural Networks (CNNs), limiting feature extraction to specific receptive fields.
• This local-to-contextual paradigm struggles with capturing global dependencies crucial for ORSI-SOD.

Purpose of the Study:

• To propose a novel Global Extraction Local Exploration Network (GeleNet) for ORSI-SOD that adopts a global-to-local paradigm.
• To enhance feature extraction by incorporating global long-range dependencies and improved local interactions.

Main Methods:

• GeleNet employs a transformer backbone for multi-level feature embeddings with global dependencies.
• It utilizes a Direction-aware Shuffle Weighted Spatial Attention Module (D-SWSAM) and its simplified version (SWSAM) for enhanced local interactions and attention.
• A Knowledge Transfer Module (KTM) facilitates cross-level contextual interactions using self-attention.

Main Results:

• GeleNet effectively captures global dependencies and enhances local feature details for salient object detection.
• The Direction-aware Shuffle Weighted Spatial Attention Module (D-SWSAM) adapts to various object orientations in ORSIs.
• Experiments on three public datasets show GeleNet surpasses state-of-the-art ORSI-SOD methods.

Conclusions:

• GeleNet offers a superior approach to ORSI-SOD by leveraging a global-to-local paradigm.
• The proposed network architecture and attention mechanisms significantly improve salient object detection performance in optical remote sensing images.
• The study provides an open-source implementation for reproducibility and further research.