Faster-CEASC+: A Lightweight Object Detection Network for Drone Images

As Unmanned Aerial Vehicle (UAV) technology continues to advance and become more widespread, object detection on UAVs has become a critical task. Constrained by the size and energy consumption of UAVs, deploying lightweight object detection models on them is a better choice. However, previous work has focused on the partial lightweighting of the network structure. This paper proposes a fully lightweight object detection model for the entire network structure, called Faster Context-Enhanced Adaptive Sparse Convolutional Plus (Faster-CEASC+). The paper's first contribution is that we design an extremely lightweight detection head, CEASC+, which maintains high accuracy. Moreover, our model also features a lightweight backbone that balances FLOPs and FPS. Extensive experiments on the VisDrone-DET2018 dataset demonstrate that compared to the state-of-the-art detector, generalized focal loss V1-CEASC, our model improves significantly by 1.1 mAP and 124.6% FPS, while reducing 60.8% FLOPs.