CAS-YOLO: A lightweight and efficient object detector for infrared UAV imagery

Abstract

Infrared target detection is crucial for unmanned aerial vehicle (UAV) applications in civilian low-visibility environments, including urban nighttime monitoring, emergency rescue, and infrastructure inspection. However, thermal imaging blur, complex backgrounds, and extreme target aspect ratios pose significant challenges to thermal infrared target detection. Moreover, conventional real-time detectors, constrained by the limited edge computing resources of UAVs, typically rely on inefficient square convolutions and computationally expensive attention mechanisms. To address these issues, we propose CAS-YOLO, an efficient and lightweight detector specifically tailored for civilian thermal infrared UAV imagery. Building upon YOLOv10n, CAS-YOLO selectively deploys Large Strip Convolution (LSC) at intermediate backbone stages to enhance the directional structures of elongated targets while reducing interference from surrounding thermal clutter. Circulant Contextual Attention (CCA) is placed at the terminal low-resolution stage to aggregate long-range contextual information without applying global modeling to high-resolution feature maps. StarReLU is applied throughout the basic convolutional units as a lightweight nonlinear transformation for weak-texture infrared features. Experiments on HIT-UAV, together with auxiliary cross-domain evaluation on FLIR, show that CAS-YOLO improves detection accuracy within a YOLOv10n-based lightweight framework. The ablation studies evaluate the individual effects of the three components and the complementary effects of LSC and CCA. This study is strictly limited to civilian applications in public safety, traffic management, and autonomous driving assistance.