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

## TL;DR

This paper introduces GMF-Net, a new lightweight and accurate method for detecting small objects in satellite laser ranging images, which improves performance while reducing computational costs.

## Contribution

The novel Gaussian-Matched Convolution module is designed based on the physical properties of SLR targets, improving detection accuracy and efficiency.

## Key findings

- GMF-Net achieves an mAP@50 of 93.1% and mAP@50–95 of 52.4% on a new SLR-CCD dataset.
- The model reduces parameters by 26.6% and computational load by 27.4% compared to baseline models.
- The GMConv module enhances feature extraction for small objects in low signal-to-noise environments.

## Abstract

Detecting small objects in Satellite Laser Ranging (SLR) CCD images is critical yet challenging due to low signal-to-noise ratios and complex backgrounds. Existing frameworks often suffer from high computational costs and insufficient feature extraction capabilities for such tiny targets. To address these issues, we propose the Gaussian-Matched Fusion Network (GMF-Net), a lightweight and high-precision detector tailored for SLR scenarios. The core scientific innovation lies in the Gaussian-Matched Convolution (GMConv) module. Unlike standard convolutions, GMConv is theoretically grounded in the physical Gaussian energy distribution of SLR targets. It employs multi-directional heterogeneous sampling to precisely match target energy decay, enhancing central feature response while suppressing background noise. Additionally, we incorporate a Cross-Stage Partial Pyramidal Convolution (CSPPC) to reduce parameter redundancy and a Cross-Feature Attention (CFA) module to bridge multi-scale features. To validate the method, we constructed the first dedicated SLR-CCD dataset. Experimental results show that GMF-Net achieves an mAP@50 of 93.1% and mAP@50–95 of 52.4%. Compared to baseline models, parameters are reduced by 26.6% (to 2.2 M) with a 27.4% reduction in computational load, demonstrating a superior balance between accuracy and efficiency for automated SLR systems.