LEGNet: A Lightweight Edge-Gaussian Network for Low-Quality Remote Sensing Image Object Detection

TL;DR

LEGNet tackles RSOD under degraded imaging by introducing Edge-Gaussian Aggregation (EGA), which combines fixed edge-sensitive (Scharr) and Gaussian-prior modules with learnable LEG blocks in a four-stage lightweight backbone. A LoG-Stem layer initiates edge-aware feature extraction, while macro designs ensure edge information is utilized in shallow layers and Gaussian priors in deeper layers, maintaining efficiency. Empirical results across DOTA-v1.0/v1.5, DIOR-R, FAIR1M-v1.0, and VisDrone2019 demonstrate state-of-the-art mAP (e.g., 80.03% on DOTA-v1.0 with LEGNet-S + O-RCNN) with substantially fewer parameters than competing backbones. The approach offers practical impact for resource-constrained RS applications, enabling robust detection of low-quality, occluded, or blurred objects with a compact model.

Abstract

Remote sensing object detection (RSOD) often suffers from degradations such as low spatial resolution, sensor noise, motion blur, and adverse illumination. These factors diminish feature distinctiveness, leading to ambiguous object representations and inadequate foreground-background separation. Existing RSOD methods exhibit limitations in robust detection of low-quality objects. To address these pressing challenges, we introduce LEGNet, a lightweight backbone network featuring a novel Edge-Gaussian Aggregation (EGA) module specifically engineered to enhance feature representation derived from low-quality remote sensing images. EGA module integrates: (a) orientation-aware Scharr filters to sharpen crucial edge details often lost in low-contrast or blurred objects, and (b) Gaussian-prior-based feature refinement to suppress noise and regularize ambiguous feature responses, enhancing foreground saliency under challenging conditions. EGA module alleviates prevalent problems in reduced contrast, structural discontinuities, and ambiguous feature responses prevalent in degraded images, effectively improving model robustness while maintaining computational efficiency. Comprehensive evaluations across five benchmarks (DOTA-v1.0, v1.5, DIOR-R, FAIR1M-v1.0, and VisDrone2019) demonstrate that LEGNet achieves state-of-the-art performance, particularly in detecting low-quality objects.The code is available at https://github.com/AeroVILab-AHU/LEGNet.

Figures (9)

• Figure 1: Visualization results on the DOTA-v1.0 test set xia2018dota. All models were built with Oriented R-CNN xie2021oriented detector. Our LEGNet demonstrates robust detection under challenging conditions such as occlusion (e.g., objects obscured by trees) and low-light (e.g., building shadows), surpassing previous state-of-the-art methods in both accuracy and robustness for low-quality objects.
• Figure 2: Overview of LEGNet architecture. It consists of 4 stages with input resolutions downsampled by factors of 4, 8, 16, and 32. AN, Conv, GAP, and $z$ denote the activation–normalization layer, convolutional layer, global average pooling, and 1D Conv size, respectively.
• Figure 3: Visualization of detection results on DOTA-v1.0 test set xia2018dota. Input images resolution were 1,024 $\times$ 1,024.
• Figure 4: Visualization of intermediate feature maps from PKINet-S (top row) and our LEGNet-S (bottom row). LEGNet's early stages capture more complete edge details, while Stage 2 demonstrates enhanced focus on salient object regions. This supports the effectiveness of our proposed design in refining feature representations for robust detection.
• Figure 5: Visualization of detection results on DOTA-v1.0 test set. Input images resolution were 1,024 $\times$ 1,024.