An Improved YOLOv8 Algorithm for Oriented Object Detection in Remote Sensing Images

Remote sensing target detection is faced with the problems of random target direction, difficult detection of small targets, and dense arrangement. In order to solve the above problems, we propose a lightweight rotating remote sensing target detection algorithm based on YOLOv8, namely Slim-YOLOR. Firstly, we reduce the number of feature map channels in the feature fusion module, and incorporate a large kernel convolutional spatial attention (LSA) module to increase the model sensing field. Secondly, dynamic upsampling is used to reduce the information loss in the process of small target feature reconstruction and enhance the detection ability of dense small targets. Finally, the shared convolution is used to interact with the classification and regression tasks, and the deformable convolution is combined to enhance the localization of the oriented target. Experimental results on the DIOR-R dataset demonstrate that compared to YOLOv8, Slim-YOLOR reduces the number of parameters by 39.5% while increasing the [email protected]:0.95 by 1.6%. Additionally, Slim-YOLOR shows clear advantages over other classical algorithms.