J. Eur. Opt. Society-Rapid Publ. 22, 43( 2026) 419
address these issues, this paper proposes the improved YOLO-KMM model, which achieves coordinated optimization of detection performance and computational efficiency through targeted module design. The main research contributions of this paper are as follows:
1) Constructed the self-built SIM-AIR dataset for air-toair UAV infrared detection, covering 4 weather conditions and ultra-small target characteristics, filling the gap of dedicated infrared datasets for air-to-air scenarios, and providing high-quality experimental data for small-target real-time detection research;
2) Proposed the C2KD feature enhancement module, |
which strengthens the model’ s ability to extract and |
represent weak features of small targets in the SIM- |
AIR dataset through cross-scale feature fusion and |
attention |
mechanisms |
[ 9 ], |
adapting |
to |
detection |
requirements |
under |
low |
SNR |
and |
complex |
backgrounds; |
|
|
|
|
|
3) Designed the C3K2-MU lightweight detection head, |
which uses grouped convolution and channel optimization |
strategies to reduce parameters and computation |
while |
ensuring |
detection |
accuracy |
[ 10 ], |
meeting |
deployment |
requirements |
in |
resource-constrained |
scenarios; |
|
|
|
4) Conducted comparative experiments with multiple mainstream YOLO models on the SIM-AIR dataset, fully verifying the comprehensive advantages of the proposed model in accuracy, speed, and computational complexity, and providing a practical technical solution for air-to-air small-target infrared detection;
5) The subsequent structure of this paper is arranged as follows: Section 2 details the construction process and feature analysis of the SIM-AIR dataset, as well as the design scheme of the YOLO-KMM model; Section 3 verifies the model’ s effectiveness through ablation experiments and performance comparison experiments; Section 4 summarizes the research results and looks forward to future directions.
2 Method
2.1 Self-built UAV air-to-air infrared dataset 2.1.1 Data acquisition
The UAV air-to-air infrared dataset used in this study is constructed based on real-scene data acquisition, aiming to address the scarcity of dedicated datasets for infrared detection of small UAVs in air-to-air scenarios.
This research utilized a multi-rotor unmanned aerial vehicle( UAV) equipped with a professional infrared thermal imaging camera to conduct air-to-air infrared data collection. The camera’ s spectral response covers the 8 – 14 lm long-wave infrared band, capable of penetrating through haze, light fog and other meteorological conditions, thereby reducing the interference of atmospheric scattering [ 11 ]. With an imaging resolution of 640 512 pixels and a pixel pitch of 12 lm, it has a strong spatial resolution capability, allowing for the clear presentation of the contours of extremely small targets even at long distances. The thermal sensitivity is 50 mk, enabling precise capture of the thermal radiation differences between targets and backgrounds. Even in low-temperature environments such as snowy days, it can distinguish the gray-scale features of both, supporting the collection of reverse contrast scenes. It supports stable imaging at 30FPS and is equipped with a three-axis mechanical gimbal anti-shake system, effectively compensating for flight attitude jitter and preventing blurring of target images. The field of view is 41.2 °– 60 ° and supports multi-level digital zoom, allowing for flexible adjustment of the imaging scale and simulation of the imaging effects of targets at different distances, ensuring the diversity and authenticity of the target size distribution in the dataset.
Acquisition Scenarios: Data are collected in open outdoor airspace and cover four typical weather conditionssunny days with a pure sky background and no atmospheric attenuation, cloudy days with uneven thermal radiation of the cloud background, snowy days with low ambient temperature and weak target-background contrast, and hazy days with high atmospheric turbidity and low SNR of infrared images-to ensure the dataset’ s diversity and practicality.
Dataset Scale: Approximately 4,000 valid infrared images are collected, all containing small UAV targets( single category: civil small UAVs with a fuselage size of 0.3 0.8 m). To ensure the fairness and reliability of model training and testing, the dataset is randomly divided into three subsets at a ratio of 7:2:1: 2,800 images for the training set, 800 for the validation set, and 400 for the test set.
2.1.2 Data Annotation
Annotation Tool: All manual annotation of infrared images is completed using the open-source image annotation tool LabelImg, which features a user-friendly graphical interface and supports direct export of YOLO-format annotation files [ 12 ].
Annotation Rules: Since the dataset contains only one target category( small UAVs), the annotation strictly follows the YOLO format specification: each image corresponds to a. txt annotation file, where each line records the target’ s category ID( small UAVs correspond to ID 0), normalized center coordinates( xcenter, ycenter), and normalized width / height. All parameters are normalized based on the 640 512 image resolution. To ensure annotation accuracy, two annotators conducted cross-validation; samples with inconsistent annotations were rechecked and corrected, resulting in a final annotation accuracy of over 99 %.
2.1.3 Dataset feature analysis
Target Size Distribution: Statistical analysis is conducted on the size of small UAV targets in infrared images, where