A Deep Neural Network with Multi-Scale Attention Mechanisms for Small Unmanned Aerial Vehicle Detection in Complex Environments

Wirtschaftsinformatik & Management · Springer Nature · 2026-11-19 · DOI: 10.1007/s10796-026-10840-1

Аннотация

In the modern transportation system, detecting small unmanned aerial vehicles (UAVs) in complex environments poses significant challenges, primarily due to their extremely small pixel proportion, presence of motion blur, and interference from cluttered backgrounds. While state-of-the-art You Only Look Once (YOLO) series detectors have achieved remarkable success in general object detection scenarios, their performance degrades significantly when confronted with tiny objects with ambiguous visual features and objects in complex backgrounds. Thus, we propose an enhanced detection architecture that integrates four attention mechanisms specifically designed to address the challenges of small UAV detection, including the Enhanced Efficient Channel Attention Module, Lightweight Convolutional Block Attention Module, Small Target Booster, and Context-Aware Attention Module. These modules are strategically positioned in the backbone and neck structures of our network, complemented by an anchor box configuration optimized for the distribution of small objects, which significantly improves the model’s performance. On a 10,000-image UAV dataset, our method attains the highest mAP50 (0.901) and recall (0.858) among all compared detectors, while remaining competitive on the stricter mAP50-95 metric. The four modules add only marginal overhead over the YOLOv5 baseline ($$+0.60$$M parameters and $$+1.39$$ GFLOPs), preserving real-time inference. Furthermore, when integrated into other YOLO backbones the proposed modules yield consistent gains on most architectures, although the magnitude of improvement is backbone-dependent. Rather than introducing a new attention primitive, our contribution is a systematic, problem-driven integration of complementary attention mechanisms tailored to small UAV detection. Code is available at https://github.com/ShinyWing/MultiScaleUAVDetection.git .

Читать у издателя →

Похожие статьи