418
J. Eur. Opt. Society-Rapid Publ. 22, 43( 2026)
Figure 1. Sample images from the dataset.
detection scenarios due to their end-to-end training mode and efficient inference performance. From YOLOv1 to the latest YOLOv11, models have continuously broken through in accuracy and speed through backbone network optimization and feature fusion mechanism upgrades [ 1 ]; however, in specific scenarios like air-to-air UAV infrared detection, two core bottlenecks remain: the lack of dedicated datasets, and the balance between small-target detection and lightweight deployment.
Air-to-air small UAV detection has distinct particularities: targets are at long distances with extremely low pixel occupancy, are highly affected by weather changes, and infrared images often suffer from low signal-to-noise ratio( SNR) and weak target-background contrast [ 2 ]. However, existing public datasets mostly focus on ground targets or conventional visual scenarios, lacking exclusive data support for air-to-air UAV infrared detection and thus failing to fully cover the complex characteristics and detection difficulties of this scenario [ 5 ]. To fill this gap, this paper constructs the self-built SIM-AIR dataset for air-to-air UAV infrared target detection, the data was collected using real-scene acquisition methods, with a drone carrying an infrared thermal imaging camera to complete the collection work. The infrared thermal imaging camera has a thermal sensitivity of 50 mk, an image resolution of 640 512 pixels, and a spectral range of 7.5 – 13.5 lm [ 3 ]. The collected dataset covers four typical weather conditions: sunny, cloudy, snowy, and smoggy. Among them, there are 2,043 sunny samples, accounting for 51.3 % of the total samples; 787 cloudy samples, accounting for 19.7 %; 620 snowy samples, accounting for 15.5 %; and 543 smoggy samples, accounting for 13.5 %. The entire dataset contains 3,993 valid images, all of which include civilian small drone targets [ 4 ]. The targets belong to a single category, with a body size ranging from 0.3 to 0.8 m. Tatistical analysis reveals that 99.7 % of the targets in the dataset are ultrasmall objects [ 6 ]. All these targets have a width of less than 40 pixels, with an average width of 11.2 pixels and an average height of 6.6 pixels, as shown in Figure 1( the dataset samples under weather conditions are presented); additionally, target SNR varies significantly across weather conditions: hazy days have an SNR as low as �0.09, while snowy days exhibit“ reverse contrast”( target grayscale lower than background) with an SNR of �0.30 [ 7 ]. These characteristics enable the SIM-AIR dataset to accurately simulate the core challenges of actual air-to-air detection, providing high-quality data support for related research.
Although newer versions like YOLOv11 and YOLOv12 have made progress through lightweight architecture design, traditional YOLO models still have obvious shortcomings when facing the SIM-AIR dataset’ s ultra-small targets, low SNR, and complex backgrounds: insufficient small-target feature extraction capability, leading to missed or false detections; meanwhile, air-to-air detection deployment platforms( e. g, UAVs, embedded devices) have limited computing resources, and existing high-precision models often come with large computational overhead [ 8 ], making it difficult to meet real-time requirements. To