LEA-DETR: A Lightweight and Efficient Attention-Enhanced Model for UAV Object Detection

Object detection in UAV scenarios has emerged as a highly significant yet challenging research topic. UAV-captured images often suffer from issues such as tiny object sizes with indistinct features, dense object distributions, and frequent occlusions. Additionally, UAV platforms are constrained by limited onboard computing power and storage capacity, which necessitates lightweight and efficient detection algorithms. In this paper, we propose LEA-DETR, a lightweight and efficient UAV object detection model based on the DETR framework. Specifically, we design an adaptive sparse-dense attention fusion module to enhance the model's ability to selectively attend to critical features. Furthermore, we introduce an efficient multi-scale convolutional FPN to effectively integrate low-level semantic information, particularly for small object detection. To further improve local representation capability, we design a novel gated inverted bottleneck convolution module, which forms the core of our newly proposed PG-Backbone for feature extraction. We conduct extensive ablation and comparison experiments on the VisDrone2019 dataset. The ablation results demonstrate that the proposed LEA-DETR-S and LEA-DETR-M models significantly reduce the number of parameters and computational overhead while achieving notable improvements in accuracy. Comparative evaluations show that LEA-DETR achieves the best overall performance among all evaluated models, making it a practical and efficient solution for UAV-based object detection in real-world scenarios.