SLIM-Net: Efficient Runway Detection via Quantization-Based Pruning

The aviation industry is increasingly adopting vision-based landing systems to improve landing safety.However, the lack of public labeled runway datasets significantly hinders the development and evaluation of robust detection systems.Although existing deep learningbased runway detection methods have shown strong performance in clear weather, their effectiveness in challenging scenarios remains limited, highlighting the need for diverse and representative datasets.Moreover, these approaches often rely on large, complex models with high computational and memory demands, making them unsuitable for real-time or on-board deployment.To address these challenges, we construct an Aerial Runway Detection Dataset (AeroRun), a benchmark featuring runway images under different weather conditions, designed to support evaluation in real-world scenarios.Also, we propose SLIM-Net, a lightweight and efficient runway detection framework that integrates standard semantic segmentation with a novel quantization-based pruning strategy.By applying bit-wise quantization and eliminating zero-valued parameters, SLIM-Net significantly reduces model size and inference cost without compromising accuracy.Experimental results on AeroRun show that SLIM-Net reduces model size by up to 29.7 and FLOPs by approximately 3.8, while maintaining detection accuracy comparable to state-of-the-art models.These findings highlight SLIM-Net's potential for real-time use in resource-limited settings and its role in improving aviation safety via reliable runway detection.