EfficientDet-EdgeUAV: A Multi-Scale Fusion Architecture for Target Detection in UAV Imagery With Edge Computing Optimization

Abstract

To overcome cloud computing's limitations—high latency and costly data transfers that hinder rapid UAV detection—plus the challenges of spotting tiny targets against complex backgrounds in wide-field aerial views, an edge computing solution with its specialized lightweight network: EfficientDet-EdgeUAV is proposed. The network employs a structurally optimized EfficientNet backbone through lightweight architecture modifications, integrating a Squeeze-and-Excitation attention mechanism to mitigate interference from complex backgrounds in target detection. The network architecture enhances small object detection by incorporating large-scale feature layers into the pyramid structure of the neck and applying lightweight architectural optimization to the neck module. The architecture further enhances detection robustness by implementing multi-scale feature fusion in the neck module, which strategically combines shallow-layer spatial details and deep-layer semantic representations to improve discernment of small objects with blurred boundaries. Through extensive experiments that comprehensively evaluate and validate the effectiveness of the proposed method, the experimental results demonstrate superior detection accuracy and efficiency on the VisDrone dataset compared to baseline and state-of-the-art methods. This demonstrates that the proposed method achieves exceptional effectiveness in real-time UAV imaging scenarios, providing critical technical references for civilian applications of drone technology in power line inspection, geological exploration, and search and rescue operations.