J. Eur. Opt. Society-Rapid Publ. 22, 43( 2026) 427
lightweight deployment strategies documented in recent embedded AI research.
4 Conclusion
Aiming at the practical challenges of air-to-air infrared small UAV detection, including the lack of dedicated datasets, sparse ultra-small target features, and limited deployment resources of on-board platforms, this study completes the construction of the SIM – AIR dataset and the design of the YOLO – KMM model, and forms a complete technical solution of“ dedicated dataset and lightweight detection model”. This section summarizes the main research results of the study, analyzes the existing limitations, and further proposes the future research directions for the optimization of the dataset and model.
The construction of the SIM – AIR dataset fills the gap of dedicated infrared datasets for air-to-air scenarios. It includes 3993 accurately annotated images, four typical weather conditions, 99.7 % ultra-small target samples, and complex scenarios such as“ reverse contrast” in snowy days and low signal-to-noise ratio( SNR) in hazy days, which fully simulates the arduousness of actual air-to-air detection. It not only provides an accurately adapted experimental platform for this study but also offers valuable highquality data support for subsequent research in related fields [ 29, 30 ]. Experimental results fully verify the practicality of the SIM – AIR dataset and the adaptability of the YOLO – KMM model: on this dataset, the mAP 50 of YOLO – KMM reaches 88.2 %, which is 7.8 % points higher than that of the baseline model; the precision and recall are increased to 94.0 % and 74.3 % respectively; the small target miss rate is significantly reduced by 12.5 %, fully proving the effectiveness of the C2KD module in enhancing weak feature extraction of small targets and suppressing background interference. Meanwhile, the model parameters are controlled at 2.3 M, the computation amount is only 5.4 GFLOPs, and the inference frame rate reaches 246.18 FPS. Compared with mainstream models such as YOLOv5, YOLOv8, YOLOv11, and YOLOv12, it shows the optimal balance performance of“ accuracy-speed-complexity”, especially in complex weather scenarios such as hazy and snowy days, with significant advantages in detection robustness.
Compared with existing research, the YOLO – KMM model, through targeted module design, is more adaptable to the ultra-small target, low SNR characteristics of the SIM – AIR dataset and the deployment requirements of resource-constrained scenarios. Its lightweight, high-precision, and fast-speed characteristics endow it with broad application prospects in practical applications such as UAV air-to-air inspection and real-time detection on embedded devices [ 31 – 34 ]. However, this study still has certain limitations: the detection performance of the model in extreme occlusion and multi-target overlap scenarios in the SIM – AIR dataset needs further improvement; at the same time, the target categories of the dataset only cover a single type of small UAV, and its generalization can be further expanded. Future research can be carried out in the following directions. First, expand the SIM-AIR dataset by systematically adding samples with extreme occlusion – defined as target obscuration exceeding 50 % of the bounding box area – along with multi-target overlap scenarios and diverse UAV platforms, which will improve the model’ s generalization capability [ 35 ]. Second, explore multi-modal sensor fusion strategies, combining early fusion approaches that integrate raw radar signals with infrared features at the feature extraction stage, and late fusion methods that merge independent infrared-based detection with radar trajectory estimation at the decision level, to substantially enhance robustness in adverse weather conditions [ 36 – 38 ]. Third, implement advanced model compression techniques including quantization and structured pruning to further reduce computational overhead for deployment on edge devices [ 39 ].
Acknowledgments
The authors would like to thank the anonymous reviewers for their valuable comments and suggestions that helped improve the quality of manuscript.
Funding
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Conflicts of interest The authors declare that they have no competing interests.
Data availability statement
The dataset generated and analyzed during the current study will be made publicly available upon acceptance of the manuscript. Author contribution statement
All authors take part in the discussion of the work described in this paper. These authors contributed equally to this work.
Ethics approval
This study does not involve human participants or animals, and therefore ethical approval is not required. Informed consent
Informedconsentisnotapplicable, asthisstudydoesnotinvolve human subjects.
References
1 Zhang T., Wu H., Liu Y., Li L., Peng J., Infrared small target detection based on local contrast measure and gradient minimization, Infrared Phys. Technol. 105, 103260( 2020). https:// doi. org / 10.1016 / j. infrared. 2020.103260.
2 Dai Y., Wu Y., Zhou F., Barnard K., A survey of infrared small target detection, IEEE Trans. Geosci. Remote Sens. 60, 1 – 15( 2022). https:// doi. org / 10.1109 / TGRS. 2022.3173446. 3 Jocher G., Chaurasia A., Qiu J., Ultralytics YOLOv8, GitHub Repository( 2023). https:// github. com / ultralytics / ultralytics.
4 Bochkovskiy A., Wang C.-Y., Liao H.-Y. M., YOLOv4: Optimal speed and accuracy of object detection, arXiv: 2004.10934( 2020). https:// doi. org / 10.48550 / arXiv. 2004.10934.