UAVLite-YOLOv8: UAV Small Target Lightweight Detection Algorithm

Aiming at the problems of UAVs in target detection, such as small target size, dense target, difficulty in feature extraction, and the excessive amount of parameters and high computation when the YOLO algorithm is applied to actual deployment, a UAV small target lightweight detection algorithm, UAVLite-YOLOv8, which is based on the improvement of YOLOv8 algorithm, is proposed. Dilation-wise Reparam Residual (DWRR) is firstly used to improve the C2f layer in the backbone, which improves the accuracy of the model while reducing the resource consumption. Global and local feature fusion (GLFF) is then used in the neck to improve the accuracy of the algorithm without significantly affecting the number of network parameters and computational effort. Finally, the detail fusion shared convolution lightweight detection (DFSCLD) head is proposed to replace the original algorithm detection head, which reduces the computational burden while maintaining the detection performance and makes the algorithm more suitable for real-time application scenarios. By conducting comparative experiments of different algorithms on the VisDrone2019 dataset, the results show that the UAVLite- YOLOv8 algorithm significantly outperforms other target detection algorithms. Compared with YOLOv8, the mAP@0.5 of UAVLite-YOLOv8 grows from 31.4% to 32.2%, the amount of parameters is reduced by 25.6%, and the amount of computation is reduced by 17.3%. In this way, it proves that the algorithm can better meet the needs of UAV small target lightweight detection.