An Improved YOLOv8 Algorithm for Rotating Object Detection in Unmanned Aerial Vehicle Remote Sensing Images

Unmanned aerial vehicle remote sensing images suffer from problems such as arbitrary object orientation and dense arrangement of small targets, which makes horizontal box object detection difficult. To address these issues, an improved YOLOv8 based remote sensing image rotating object detection algorithm (R-YOLOv8) is proposed. Firstly, the long edge definition method of the five-parameter method is used to represent the rotation box, which allows the bounding box to cover the small target at any angle. This can better fit the actual shape and arrangement of the target, which effectively solves the problem of accidental deletion caused by inaccurate horizontal box positioning. Then, ProbIOU Loss is introduced as the regression loss function to convert the rotation box into a two-dimensional Gaussian distribution, which solves the problem of angle periodicity caused by the representation of rotating targets. Finally, the BiFormer attention mechanism is introduced to improve the accuracy of target detection while reducing parameters and computational complexity. This paper conducts relevant experiments on the UAV-ROD dataset. Compared to classical networks and existing oriented object detection networks, the R-YOLOv8 algorithm performs better in the case of a model size of only 2.8 MB, [email protected] reaches 98.7%, which verifies the effectiveness and progressiveness of the proposed method.