YOLO-LE: A lightweight and efficient UAV aerial image target detection model

Abstract

We designed a lightweight and efficient target detection algorithm YOLO-LE: 1) By designing the C2f-Dy and LDown modules, the small target feature sensitivity of the backbone is improved, while the number of backbone parameters is reduced and the model efficiency is improved. 2) By designing an adaptive feature fusion module, we can flexibly integrate feature maps of different sizes, optimize the neck architecture, lightweight the neck network, and improve model performance. 3) We replace the loss function of the original model with a distributed focal loss and combine it with a simple self-attention mechanism by design to improve small object recognition and anchor box regression performance.In comparative experiments on the VisDrone2019 dataset, our YOLO-LE model improves mAP(0.5) by 9.6% compared to YOLOv8n.The results show that our method can effectively improve the model performance.