LODNU: lightweight object detection network in UAV vision

With the development of unmanned aerial vehicle (UAV) technology, using UAV to detect objects has become a research focus. However, most of the current object detection algorithms based on deep learning are resource consuming, which is difficult to deploy on embedded devices lacking memory and computing power, such as UAV. To meet these challenges, this paper proposes a lightweight object detection network in UAV vision (LODNU) based on YOLOv4, which can meet the application requirements of resource-constrained devices while ensuring the detection accuracy. Based on YOLOv4, LODNU uses depth-wise separable convolution to reconstruct the backbone network to reduce the parameters of the model, and embeds improved coordinate attention in the backbone network to improve the extraction ability of key object features. Meanwhile, the adaptive scale weighted feature fusion module is added to the path aggregation network to improve the accuracy of multi-scale object detection. In addition, in order to balance the proportion of sample size, we propose a patching data augmentation. LODNU achieves the performance near to YOLOv4 with fewer parameters. Compared with YOLOv4, LODNU achieves more balanced results between model size and accuracy in the VisDrone2019 dataset. The experimental show that the parameters of LODNU are only 13.1% of YOLOv4, and the floating-point operations are only 13.6% of YOLOv4. The mean average precision of LODNU in the VisDrone2019 dataset is 31.4%, which is 77% of YOLOv4.