SA-YOLO: Spike-Driven Attention for Energy-Efficient UAV-Based Small Object Detection

Unmanned Aerial Vehicles (UAVs) are increasingly employed in Internet of Things (IoT) applications to enable real-time visual perception. However, the limited computational and energy resources available onboard demand low-latency and energy-efficient edge processing. Moreover, the low resolution and scale variation of aerial imagery present significant challenges for accurate object detection. To address these challenges, we propose the Spike Multi-Scale Attention YOLO (SA-YOLO), a spiking neural network (SNN)-based object detection model designed for UAV scenarios, which achieves a favorable balance between detection accuracy and energy efficiency. SA-YOLO leverages event-driven, sparse computations that primarily involve addition, significantly reducing energy consumption compared to ANN-based methods that rely heavily on multiplication operations. To further improve efficiency, a multi-scale attention mechanism (S-MSA) is proposed, which integrates temporal and channel-wise features to enhance informative feature extraction while reducing spike firing rates. Unlike ANN-to-SNN conversion approaches, which transfer pre-trained artificial neural networks to the spiking domain and typically require many time steps, the proposed method adopts direct training with spiking neurons and achieves comparable recognition performance using only two time steps, leading to improved energy efficiency and lower computational cost. Experimental results on HIT-UAV and CARPK datasets demonstrate mAP@0.5 of 87.33% and 95.60% with improved energy efficiency. Our model surpasses mainstream lightweight SNN models in both detection accuracy and computational efficiency, while ensuring robust performance across diverse UAV environments.