A Lightweight Spiking Neural Network for Real-Time Object Recognition on UAV Edge Platforms

Real-time target recognition on unmanned aerial vehicles (UAVs) requires algorithms that simultaneously deliver high accuracy and low power consumption under strict resource constraints. In this paper, we propose a lightweight spiking neural network (SNN) framework for electro-optical image recognition on UAV platforms. To bridge the gap between frame-based sensing and event-driven processing, an encoding module is introduced to efficiently transform static images into temporal spike trains. For deep feature representation, we design a spiking residual network that incorporates membrane potential residual connections, thereby enhancing information transmission. Furthermore, convolutional layers are integrated with batch normalization to reduce computational overhead during inference while maintaining accuracy. Experimental results on the Skyview aerial dataset demonstrate that the proposed Spiking-ResNet18 and Spiking-ResNet34 achieve recognition accuracies of 96.6% and 97.1%, respectively, comparable to their conventional ANN counterparts. More importantly, the energy consumption of our models is reduced to only 18.6% and 15.9% of equivalent ANN architectures, highlighting their superior energy efficiency. These results validate the practicality and high efficiency of the proposed algorithm, underscoring the potential of SNN-based architectures for real-time, low-power target recognition on resource-constrained UAV platforms.