SLR-Net: Lightweight and Accurate Detection of Weak Small Objects in Satellite Laser Ranging Imagery

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Traditional detection models struggle with minute targets, low signal-to-noise ratios (SNRs), and feature volatility in Satellite Laser Ranging (SLR) images.
• Existing methods often face challenges in efficiency and accuracy due to these limitations.

Purpose of the Study:

• To propose an efficient, lightweight, and high-precision detection model for minute and blurred targets in SLR images.
• To enhance feature extraction, fusion, and localization capabilities without a significant computational burden.

Main Methods:

• Designed a Dense Multi-Scale Convolution (DMS-Conv) module for expanded receptive fields and improved feature representation of faint targets.
• Introduced a Lightweight Upsampling Module (LUM) using depthwise separable convolutions for efficient multi-scale feature fusion.
• Developed a Modified Progressive Diagonal Intersection over Union (MPD-IoU) Loss function for precise small target localization.

Main Results:

• The proposed model achieved an mAP50:95 of 47.13% and an F1-score of 88.24% on a real-world SLR dataset.
• The model has only 2.57 million parameters and 6.7 GFLOPs, demonstrating a lightweight design.
• Outperformed various mainstream lightweight detectors in precision and recall.

Conclusions:

• The developed model effectively addresses small target detection challenges in SLR scenarios.
• The lightweight design and superior performance offer significant practical value for SLR image analysis.
• The proposed DMS-Conv, LUM, and MPD-IoU Loss contribute to enhanced feature representation, fusion, and localization.