Lightweight Multi-Scale AI for Small Aerial Target Recognition Using Micro-Doppler Signatures
Year: 2026
Type: conference-paper
Abstract: To mitigate security threats from small unmanned aerial vehicles (UAVs), this research proposes a lightweight convolutional neural network (CNN) architecture leveraging micro-Doppler signatures (MDS) to distinguish drones from birds. The proposed model employs multi-scale parallel convolutional branches and adaptive image compression to minimize computational costs while maintaining high recognition accuracy. Experimental results demonstrate that the architecture achieves a peak accuracy of 99.48% at$224 \times 224$resolution. For edge deployment, the model utilizes$64 \times 64$resolution to maintain a robust 97.84% accuracy with only 23.1K parameters and 4.32M FLOPs. By optimizing feature extraction and model design, this work provides an efficient and practical solution for real-world UAV identification on resource-constrained systems.
Institution: National Defense University
Language: English
Subfield: Aerospace Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204118912