Scatter-aware interaction network for robust SAR object detection

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Synthetic aperture radar (SAR) object detection faces challenges in complex environments like harbors due to speckle noise and clutter.
• These issues cause missed detections, false alarms, and unstable object localization.

Purpose of the Study:

• To develop an advanced object detection framework for SAR imagery in complex backgrounds.
• To improve accuracy, reduce false alarms, and enhance localization stability for SAR object detection.

Main Methods:

• Proposed Scatter-Aware Interaction Network (SAI-Net) with three components: Shift-wise Conv (SWC) backbone, BiFPN-Rep Residual (BRR) Fusion neck, and Enhanced AIFI (E-AIFI).
• SWC backbone for multi-scale feature extraction with enlarged receptive fields.
• BRR Fusion neck for stable multi-scale feature fusion with cross-scale alignment.
• E-AIFI for scatter-aware interactive encoding, including self-attention and feed-forward modules.

Main Results:

• SAI-Net demonstrated consistent improvements in overall accuracy on OGSOD and HRSID datasets.
• Significant enhancements were observed in small-object detection and strict localization metrics.
• The proposed scatter-aware interaction and cross-scale fusion effectively addressed challenges in complex SAR scenarios.

Conclusions:

• SAI-Net offers an effective end-to-end solution for SAR object detection in challenging environments.
• The framework's design successfully mitigates the impact of speckle noise and clutter.
• SAI-Net provides a robust approach for accurate and stable object detection in complex SAR imagery.