Improved high-precision UAV small object detection model based on YOLOv8

Object detection for unmanned aerial vehicle (UAV) has always been a popular research direction. However, the detection often involves the issue of objects with small pixel values. In response to this issue, this paper hopes to have a detection model that can achieve precise detection of UAV small objects, while also requiring the characteristic of occupying fewer resources. Therefore, this paper proposed an improved model based on YOLOv8 [1], high-precision YOLOv8, which has good performance in UAV small object detection and has been optimized for the balance between performance and resources. Firstly, this paper made improvements to the backbone network of YOLOv8 and proposed a new backbone network; In addition, for the detection of UAV small objects, this paper designed new feature fusion modules and a new detection head. Finally, this paper used MPDIoU [2] as the bounding box loss, adopting the shortest point distance method to compare any two bounding boxes, considering more computational factors and avoiding the failure of general intersection over union (IoU). At the same time, this paper used the LAMP [3] pruning method to simplify the model, reduce computational complexity, and thus improve the model's localization ability. On the VisDrone2021 [4] dataset, our model showed a 71% reduction in parameter count, an 8.1% increase in mAP50, and a 5.3% increase in mAP50-95 compared to the baseline model. Compared with some excellent detection networks, the model proposed in this paper achieved considerable speed, low parameters, and best accuracy.