LIRS-YOLO: A Lightweight YOLO-Based Method for Infrared Features and Small Object Detection

Target detection in UAV-based infrared imagery plays a critical role in personnel search and rescue, security surveillance, and disaster monitoring. However, infrared targets captured from UAV platforms are often small, low-contrast, and weakly textured, and they exhibit significant scale variations due to constraints on flight altitude and viewing perspective. These challenges hinder accurate feature extraction and degrade detection performance, severely limiting the effectiveness of existing detectors. Furthermore, the limited computational capability and memory of UAV hardware make it difficult to meet real-time processing requirements.To address these issues, we propose LIRS-YOLO, a novel lightweight infrared target detection framework. Specifically, LIRS-YOLO adopts GHGNetv2 as the backbone to efficiently extract features with reduced parameters and computation. Moreover, we design a plug-and-play Deformable Thermal–Contrast Spatial Attention module (DTCSA), which leverages local thermal-contrast-inspired differences to enhance regions of interest and improve target localization. In addition, an efficient multi-scale detection head is developed to strengthen contextual reasoning and improve detection performance for targets at different scales. Extensive experiments on the HIT-UAV and DroneVehicle datasets demonstrate the superiority of LIRS-YOLO. On the HIT-UAV dataset, LIRS-YOLO outperforms state-of-the-art methods by at least 0.7% in mAP₅₀. On the DroneVehicle dataset, it achieves state-of-the-art performance (83.4% mAP₅₀). For small object, AP_s and AR_s are improved by 2.1% and 3.6%, respectively. These results validate the robustness and efficiency of LIRS-YOLO across different target types and datasets. Code is available at https://github.com/Samjiu/LIRS-YOLO.