GMF-Net: A Gaussian-Matched Fusion Network for Weak Small Object Detection in Satellite Laser Ranging Imagery.

Area of Science:

• Remote Sensing
• Computer Vision
• Signal Processing

Background:

• Satellite Laser Ranging (SLR) image analysis faces challenges with small object detection due to low signal-to-noise ratios and complex backgrounds.
• Current detection frameworks exhibit high computational costs and inadequate feature extraction for minute targets.

Purpose of the Study:

• To develop a lightweight, high-precision object detector for SLR scenarios, addressing limitations of existing methods.
• To introduce a novel Gaussian-Matched Convolution (GMConv) module for enhanced feature extraction and noise suppression.

Main Methods:

• Proposed the Gaussian-Matched Fusion Network (GMF-Net) featuring GMConv for precise target energy matching.
• Incorporated Cross-Stage Partial Pyramidal Convolution (CSPPC) to minimize parameters and Cross-Feature Attention (CFA) for multi-scale feature integration.
• Constructed the first dedicated SLR-CCD dataset for method validation.

Main Results:

• GMF-Net achieved mAP@50 of 93.1% and mAP@50-95 of 52.4% on the SLR-CCD dataset.
• Reduced model parameters by 26.6% (to 2.2 M) and computational load by 27.4% compared to baseline models.
• Demonstrated a superior balance between detection accuracy and computational efficiency.

Conclusions:

• GMF-Net offers a significant advancement in automated SLR systems for small object detection.
• The proposed GMConv module effectively enhances central feature responses while mitigating background noise.
• The GMF-Net provides an efficient and accurate solution for critical SLR applications.